Plaque Imaging: Pixel to Molecular Level (Studies in Health Technology and Informatics, Vol. 113) (Studies in Health Technology and Informatics)

PLAQUE IMAGING: PIXEL TO MOLECULAR LEVEL Studies in Health Technology and Informatics This book series was started in ...

Author: J.S. Suri; C. Yuan; D.L. Wilson; S. Laxminarayan

18 downloads 955 Views 16MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

PLAQUE IMAGING: PIXEL TO MOLECULAR LEVEL

Studies in Health Technology and Informatics This book series was started in 1990 to promote research conducted under the auspices of the EC programmes Advanced Informatics in Medicine (AIM) and Biomedical and Health Research (BHR), bioengineering branch. A driving aspect of international health informatics is that telecommunication technology, rehabilitative technology, intelligent home technology and many other components are moving together and form one integrated world of information and communication media. The complete series has been accepted in Medline. In the future, the SHTI series will be available online. Series Editors: Dr. J.P. Christensen, Prof. G. de Moor, Prof. A. Hasman, Prof. L. Hunter, Dr. I. Iakovidis, Dr. Z. Kolitsi, Dr. Olivier Le Dour, Dr. Andreas Lymberis, Dr. Peter Niederer, Prof. A. Pedotti, Prof. O. Rienhoff, Prof. F.H. Roger France, Dr. N. Rossing, Prof. N. Saranummi, Dr. E.R. Siegel and Dr. Petra Wilson

Volume 113 Recently published in this series Vol. 112. T. Solomonides, R. McClatchey, V. Breton, Y. Legré, S. Nørager (Eds.), From Grid to Healthgrid Vol. 111. J.D. Westwood, R.S. Haluck, H.M. Hoffman, G.T. Mogel, R. Phillips, R.A. Robb, K.G. Vosburgh (Eds.), Medicine Meets Virtual Reality 13 Vol. 110. F.H. Roger France, E. De Clercq, G. De Moor and J. van der Lei (Eds.), Health Continuum and Data Exchange in Belgium and in the Netherlands – Proceedings of Medical Informatics Congress (MIC 2004) & 5th Belgian e-Health Conference Vol. 109. E.J.S. Hovenga and J. Mantas (Eds.), Global Health Informatics Education Vol. 108. A. Lymberis and D. de Rossi (Eds.), Wearable eHealth Systems for Personalised Health Management – State of the Art and Future Challenges Vol. 107. M. Fieschi, E. Coiera and Y.-C.J. Li (Eds.), MEDINFO 2004 – Proceedings of the 11th World Congress on Medical Informatics Vol. 106. G. Demiris (Ed.), e-Health: Current Status and Future Trends Vol. 105. M. Duplaga, K. Zieliński and D. Ingram (Eds.), Transformation of Healthcare with Information Technologies Vol. 104. R. Latifi (Ed.), Establishing Telemedicine in Developing Countries: From Inception to Implementation Vol. 103. L. Bos, S. Laxminarayan and A. Marsh (Eds.), Medical and Care Compunetics 1 Vol. 102. D.M. Pisanelli (Ed.), Ontologies in Medicine Vol. 101. K. Kaiser, S. Miksch and S.W. Tu (Eds.), Computer-based Support for Clinical Guidelines and Protocols – Proceedings of the Symposium on Computerized Guidelines and Protocols (CGP 2004) ISSN 0926-9630

Plaque Imaging: Pixel to Molecular Level

Edited by

Jasjit S. Suri Fischer Imaging Corporation

Chun Yuan University of Washington, Seattle, USA

David L. Wilson Case Western Reserve University, Cleveland, USA

and

Swamy Laxminarayan Idaho State University, Pocatello, USA

Amsterdam • Berlin • Oxford • Tokyo • Washington, DC

© 2005 The authors. All rights reserved All rights reserved. No part of this book may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without prior written permission from the publisher. ISBN 1-58603-516-9 Library of Congress Control Number: 2005925943 Publisher IOS Press Nieuwe Hemweg 6B 1013 BG Amsterdam Netherlands fax: +31 20 620 3419 e-mail: [email protected] Distributor in the UK and Ireland IOS Press/Lavis Marketing 73 Lime Walk Headington Oxford OX3 7AD England fax: +44 1865 750079

Distributor in the USA and Canada IOS Press, Inc. 4502 Rachael Manor Drive Fairfax, VA 22032 USA fax: +1 703 323 3668 e-mail: [email protected]

LEGAL NOTICE The publisher is not responsible for the use which might be made of the following information. PRINTED IN THE NETHERLANDS

Dedication

Jasjit Suri would like to dedicate this handbook to Swamy Laxminarayan, who has dedicated his life for the growth of Biomedical Engineering.

Chun Yuan would like to dedicate this handbook to his family and students.

David Wilson would like to dedicate this handbook to his family and students.

Swamy Laxminarayan would like to dedicate this book in memory of his beloved parents who were a constant source of inspiration in his life and to his in-laws Corie and Derk Zwakman for their genuine sense of family attachments and friendship.

This page intentionally left blank

Plaque Imaging: Pixel to Molecular Level J.S. Suri et al. (Eds.) IOS Press, 2005 © 2005 The authors. All rights reserved.

vii

Preface

Chapter 1 presents the review of the plaque imaging techniques. It introduces the segmentation techniques for plaque classification and quantification. Chapter 2 we present an overview of the work on medical image retrieval and present a general framework of medical image retrieval based on plaque appearance. We stress on two basic features of medical image retrieval based on plaque appearance: plaque medical images contain complex information requiring not only local and global descriptors but also context determined by image features and their spatial relations. Additionally, given that most objects in medical images usually have high intra- and inter-patient shape variance, retrieval based on plaque should be invariant to a family of transformations predetermined by the application domain. To illustrate the medical image retrieval based on plaque appearance, we consider a specific image modality: intravascular ultrasound images and present extensive results on the retrieval performance. The increasing amount of medical images produced and stored daily in hospitals needs a database management system to organizes them in a meaningful way, without the necessity of time-consuming textual annotations for each image. One of the basic ways to organize medical images in taxonomies consists of clustering them depending of plaque appearance (for example, intravascular ultrasound images). Although lately, there has been a lot of research in the field of Content-Based Image Retrieval systems, mostly these systems are designed for dealing a wide range of images but not medical images. Medical image retrieval by content is still an emerging field, and few works are presented in spite of the obvious applications and the complexity of the images demanding research studies. Chapter 3 reviews current MRI techniques to differentiate stable versus high risk atherosclerosis and discusses the development of non-invasive MR imaging techniques to characterize atherosclerotic plaques. Tissue specific MR signal features will be described according to histo-pathological evaluation standards and comprehensive imaging protocol for the identification of different lesion types will be introduced. Chapter 4 presents the material composition and morphology of the atherosclerotic plaque as these components are considered to be more important determinants of acute coronary ischemic syndromes than the degree of stenosis. When a vulnerable plaque ruptures it causes an acute thrombotic reaction. Rupture prone plaques contain a large lipid pool covered by a thin fibrous cap. The stress in these caps increases with decreasing thickness. Additionally, the cap may be weakened by macrophage infiltration. IntraVascular UltraSound (IVUS) elastography might be an ideal technique to assess the presence of lipid pools and to identify high stress regions. Elastography is a technique that assesses the local elasticity (strain and modulus) of tissue. It is based on the principle that the deformation of tissue by a mechanical excitation is a function of its material properties. The deformation of the tissue is determined using ultrasound. For intravascu-

viii

lar purposes, the intraluminal pressure is used as the excitation force. The radial strain in the tissue is obtained by cross-correlation techniques on the radio frequency signals. The strain is color-coded and plotted as a complimentary image to the IVUS echogram. IVUS elastography, and IVUS palpography (which uses the same principle but is faster and more robust), have been extensively validated using simulations and by performing experiments in vitro and in vivo with diseased arteries from animals and humans. Strain was shown to be significantly different in various plaque types (absent, fatty, fibrous or calcified). A high strain region with adjacent low strain at the lumen vessel-wall boundary has 88% sensitivity and 89% specificity for detecting vulnerable plaques. High strain regions at the lumen plaque-surface have 92% sensitivity and 92% specificity for identifying macrophages. Furthermore, the incidence of vulnerable-plaque-specific strain patterns in humans has been related to clinical presentation (stable angina, unstable angina or acute myocardial infarction) and the level of C-reactive protein. In conclusion, the results obtained with IVUS (strain and modulus) elastography/palpography, show the potential of the technique to become a unique tool for clinicians to assess the vulnerability and material composition of plaques. Chapter 4 presents the analysis of the soft tissue based on computer vision techniques. Numerical simulations, which are based on reliable biomechanical models of blood vessels, can help to get a better understanding of cardiovascular diseases such as atherosclerosis, and can be used to develop optimal medical treatment strategies. Blood vessels consist of three different soft tissue layers that all have different mechanical properties. The adventitia (tunica externa) is the outer most layer and its mechanical properties are essentially determined by the three-dimensional, structural arrangement of collagen fibre bundles embedded in the tissue. Global information such as the orientation statistics of the fibre bundles as well as detailed information as the crimp of the single fibres within the bundles is of particular interest in biomechanical modeling. In order to obtain a sufficiently large amount of data for biomechanical modeling, a fully automatic method for the structural analysis of the soft tissue is required. In this contribution we present methods based on computer vision to fulfill this task. We start by discussing proper tissue preparation and imaging techniques that have to be used to obtain data, which reliably represents the real three-dimensional tissue structure. The next step is concerned with algorithms that robustly segment the collagen fibre bundles and cope with various kinds of artifacts. Due to the wide variety of different appearances of collagen fibres in images the segmentation is non trivial. Novel segmentation techniques for robust segmentation of individual fibril bundles and methods for estimation of their parameters, such as location, shape, fibril density, mean fibril orientation, crimp of fibrils, etc. is discussed. The proposed algorithms are based on novel perceptual grouping methods operating on the extracted orientation data of fibrils. Finally, we demonstrate the results obtained by our fully automatic method on real data. In addition, for a more quantitative assessment, we introduce a generative structural model that enables the synthesis of three-dimensional fibre bundles with well-defined characteristics. Chapter 6 we present recent developments in the modeling of coronary artery biomechanics. We first introduce the pathology and localization of lesions in the circulatory system. Recent fluid and structural modeling of CAD is presented and discussed. At the end of the chapter, we present recent effort in coupling these two modeling domains using fluid-structure interaction (FSI).

ix

Chapter 7 presents the Cardiac CT for the assessment of cardiovascular pathology with an emphasis on the detection of coronary atherosclerosis. Cardiac CT is a robust technology for the non-invasive assessment for a spectrum of cardiovascular disease processes. This imaging modality can provide assessment of atherosclerotic plaque burden and coronary artery disease risk through coronary calcium scoring. Advances in spatial and temporal resolution, electrocardiographic triggering methodology, and image reconstruction software have helped in the evaluation of coronary artery anatomy and vessel patency, providing the ability to noninvasively diagnose or rule out significant epicardial coronary artery disease. This technique also allows the 3-dimensional simultaneous imaging of additional cardiac structures including coronary veins, pulmonary veins, atria, ventricles, aorta and thoracic arterial and venous structures, with definition of their spatial relationships for the comprehensive assessment of a variety of cardiovascular disease processes. Chapter 8 details technical and practical issues regarding coronary atherosclerotic plaque imaging by CT, which then help define its technical capabilities and engineering limitations for clinical usage. The focus will be on coronary calcium imaging, but the principles are the same for atherosclerosis definition in any major artery. Chapter 9 demonstrates the feasibility of a non-invasive, in vivo determination of the IBS of arterial wall tissues, such as atherosclerotic plaques, despite the phase aberration of the intervening tissues. Studies done on carotid arteries suggest that the morphology and composition of atherosclerotic plaque are predictive of stroke risk. The goal of this investigation has been to demonstrate that the true acoustic integrated backscatter (IBS) from plaque regions can be measured non-invasively, based on which plaque composition may be inferred and thus become a tool to estimate the likelihood of a lesion or plaque being stable or vulnerable, i.e. having a risk of causing a stroke. To obtain the true IBS non-invasively, the scattering and aberrating effect of the intervening tissue layers must be overcome. This is achieved by using the IBS from arterial blood as a reference backscatter, specifically the backscatter from a blood volume along the same scan line as and adjacent to the region of interest. We have shown that the variance of the IBS estimate of the blood backscatter signal can be quantified and reduced to a specified tolerable level. Stroke is the third leading cause of death in the western world and the major cause of disability in adults. The objective of this work was to develop a computer aided system that will facilitate the automated characterization of carotid plaques for the identification of individuals with asymptomatic carotid stenosis at risk of stroke. Ultrasound scans of carotid plaques were performed using duplex scanning and color flow imaging. From the segmented plaque images texture feature sets and shape parameters were extracted. The plaques were classified into two types: (i) symptomatic because of ipsilateral hemispheric symptoms, or (ii) asymptomatic because they were not connected with ipsilateral hemispheric events. For the plaque classification a modular system composed of neural self-organizing feature map (SOM) classifiers or statistical K-nearest nearest classifiers, was used. For each feature set, a classifier was trained and their classification results were combined using majority voting and weighted averaging. In Chapter 10 two independent studies were conducted. In the first study with 230 plaques, ten different texture feature sets were extracted and used for classification. The ten classification results were further combined in order to improve the diagnostic yield. In the second study with 330 plaques, morphological features were extracted and their

x

classification results were compared with the results obtained by the most successful texture feature sets of the first study. Both studies yielded comparable results verifying each other’s correctness. The highest diagnostic yield was 73.1% and was achieved by combing the ten SOM classifiers using weighted averaging. In conclusion, the results in this work show that it is possible to identify a group of patients at risk of stroke using neural network technology and texture features extracted from high resolution ultrasound images of carotid plaques. This group of patients may benefit from a carotid endarterectomy whereas other patients may be spared from an unnecessary operation. Chapter 11 addresses the problem of reliable feature extraction and classification process. Derivatives of Gaussian filter’s bank, co-occurrence matrices measures, cumulative moments and local binary patterns are the feature extraction processes selected for comparison. The classification process used is the highly discriminative, adaptative boosting (AdaBoost). As a result, the recognition rate of each different pair of plaques involved in IVUS plaque characterization is assessed in an objective IVUS tissue database. It is known that is there is a high correlation between plaque rupture or endothelial erosion with subsequent thrombosis formation and acute coronary syndromes. The highrisk plaque is extremely related to the composition and morphology of the plaque itself. Therefore, plaque characterization in IVUS images is of great importance to the medical community since they provide a feasible way for identification of high-risk plaques. The identification of the different kind of plaques in medical imaging requires basically of two steps: First, a reliable feature extraction process, that characterizes the different plaques to be distinguished. And second, a classification process that labels each incoming new plaque in one of the desired classes. Chapter 12 presents advanced mathematical techniques to extract spectral information from the RF data to determine the plaque composition. IVUS is a minimally invasive imaging modality that provides cross-section images of arteries in real-time, allowing visualization of atherosclerosis plaques in vivo. In standard IUVS gray scale images, calcified region of plaque and dense fibrous components generally reflect ultrasound ultrasound energy well and thus appear bright and homogeneous IVUS images. Conversely regions of low echo reflectance in IVUS images are usually is labeled as soft or mixed plaque. However this visual interpretation has been demonstrated to be very inconsistent in accurately determining plaque composition and does not allow real time assessment of qualitative plaque constituents. Spectral analysis of the backscattered RF ultrasound signals allows detailed assessment of plaque composition. Advanced mathematical techniques can be employed to extract spectral information from these RF data to determine composition. The spectral content or signature of RF data reflected from tissue depends on density, compressibility, concentration and size, etc. A combination of spectral parameters were used to develop the statistical classification scheme of analysis of in-vivo IVUS data in real-time. The clinical data acquisition system in ECG gated and the analysis software developed by our group reconstructs IVUS images from from acquired RF data. A combination of spectral parameters and active contour models is used for 3-D of plaque segmentation followed by color-coded tissue maps for each image cross-section and longitudinal views of the entire vessel. The “fly-through” mode allows one to visualize the complete length of the artery internally with the histology components at the lumen surface. In addition, the vessel and plaque matrices such as areas and

xi

volumes of individual plaque components (collagen, fibro-lipid, calcium, lipid-core) are also available. Angiography or intravascular ultrasound (IVUS). Angiography provides information about the vessel lumen and its geometry. IVUS offers more detailed information that also includes the vascular wall. Chapter 13 describes these two imaging modalities and their geometrically correct fusion yielding a 3-D and/or 4-D representation of the coronary geometry and morphology. The image-derived information is used for assessment of coronary function and plaque severity, blood flow related indices are determined using computational fluid dynamics. Detailed description of the methodology is followed by validation and clinical studies. Chapter 14 reviews the state of the art in contrast-enhanced MRI of atherosclerosis. It begins with a section describing observed late-phase enhancement characteristics and their association with tissue types. The bulk of the chapter will discuss the challenges and quantitative advantages of dynamic contrast-enhanced imaging of plaque. Then a brief overview will be presented of tissue-specific agents such as USPIOs that target specific plaque features. These topics are illustrated with a variety of case studies. In all cases, subjects provided informed consent and the studies were approved by the institutional review boards. The chapter will close with a discussion of the comparative merits of each of these techniques. Chapter 15 presents the research to test the hypotheses that (1) vessel wall volume measurements from dark blood MR images with multiple contrast-weightings (T1W, T2W and PDW) are highly reproducible, and that (2) the intra-observer and interobserver variability of carotid wall volume measurements will be less than those obtained with maximum wall area (MaxWA) measurements. Methods: Sixteen patients (aged 72 ± 7 years) with carotid stenosis documented by duplex ultrasound were recruited for the study. Dark blood T1W, PDW and T2W MR images were used to measure carotid wall volume and MaxWA by two independent observers for inter-observer and intra-observer variability assessment. Results: The intra-observer absolute difference of carotid wall volume for T1W, T2W and PDW images were 67.3 ± 47.5 mm3 (2.3 ± 1.8%), 63.2 ± 52.2 mm3 (2.0 ± 1.3%), and 69.8 ± 45.2 mm3 (2.4 ± 1.7%) respectively. The inter-observer absolute difference of carotid wall volume for T1W, T2W and PDW images were 103.5 ± 141.8 mm3 (3.0 ± 3.1%), 95.9 ± 102.1 mm3 (3.1 ± 2.6%), and 132.1 ± 87.8 mm3 (4.3 ± 2.7%) respectively. The intra-observer absolute difference of carotid MaxWA for T1W, T2W and PDW images were 6.9 ± 5.0 mm2 (4.2 ± 2.9%), 5.1 ± 4.2 mm2 (3.1 ± 2.3 %) and 7.5 ± 4.7 mm2 (4.2 ± 2.7 %) respectively. The interobserver absolute difference of carotid MaxWA for T1W, T2W and PDW images were 9.5 ± 4.2 mm2 (5.8 ± 2.3%), 6.4 ± 6.1 mm2 (3.8 ± 3.1 %) and 10.8 ± 7.3 mm2 (6.1 ± 3.7 %) respectively. Both intra- and inter-observer variability in carotid volume measurement tend to be smaller than that in carotid MaxWA measurement with intraclass correlation coefficients ranged 0.932 to 0.987 for volume measurement and 0.822 to 0.946 for MaxWA measurement. Chapter 16 presents a three-dimensional image registration algorithm for magnetic resonance (MR) images of carotid vessels. We used a mutual information registration algorithm to compensate movements between image acquisitions. Proton density (PD), T1, and T2 images were acquired from patients and volunteers and then matched for image analysis. Visualization methods such as contour overlap showed that vessels well aligned after registration. Distance measurements from the landmarks indicated that the

xii

registration method worked well with an error of 1.09 ± 0.42 mm. Potential applications include atherosclerotic plaque characterization and plaque burden quantification vectorbased segmentation using dark blood MR images having multiple contrast weightings (PD, T1, and T2). Another application is measurement of disease progression and regression with drug trials. Chapter 17 will present techniques for characterizing blood flow patterns in large arteries from magnetic resonance angiography (MRA) and velocity-encoded phasecontrast magnetic resonance imaging. Considerable evidence has emerged that disturbed blood flow patterns are a major factor in the onset of atherosclerotic disease and may play a role in disease progression. This technique, known as vascular computational fluid dynamics (CFD), has been applied extensively to the bifurcation region of the carotid artery, a common site of plaque formation. Common hemodynamic features in this region will be presented based on imaging of a series of normal subjects. Hemodynamic features in the vicinity of the carotid bifurcation will also be presented for a series of subjects with advanced atherosclerotic disease. Chapter 18 will discuss some of the modeling tools developed to study drug delivery by drug eluting stents. In order to successfully prevent restenosis, drug eluting stents must deliver a therapeutic drug dose evenly through the treatment region for a defined duration. Favorable results have come from animal trials in porcine models, unfortunately information about the dose distribution over time is difficult to obtain from investigations of this nature. Moreover, porcine models of drug-eluting stents concentrate on measuring the degree of restenosis rather than drug concentrations in the tissue of interest. Clinical studies have shown impressive results with sirolimus-eluting stents even in complex disease vessels, however failure of the target vessel still occur in up to 10% in the first year.

xiii

The Editors Dr. Jasjit S. Suri received his BS in Computer Engineering in distinction from Maulana Azad College of Technology, Bhopal, India, his MS in Computer Sciences from University of Illinois, Chicago, and Ph.D. in Electrical Engineering from University of Washington, Seattle. He has been working in the field of Computer Engineering/Imaging Sciences for over 20 years. He has published more than 125 technical papers in body imaging. He is a lifetime member of research engineering societies: Tau-Beta-Pi, Eta-Kappa-Nu, Sigma-Xi and a member of NY Academy of Sciences, Engineering in Medicine and Biology Society (EMBS), SPIE, ACM and is also a Senior Member of IEEE. He is in the editorial board/reviewer of several international journals such as: Real Time Imaging, Pattern Analysis and Applications, Engineering in Medicine and Biology, Radiology, Journal of Computer Assisted Tomography, IEEE Transactions of Information Technology in Biomedicine and IASTED Board. He has chaired image processing tracks at several international conferences and has given more than 40 international presentations/seminars. Dr. Suri has written six books in the area of body imaging (such as Cardiology, Neurology, Pathology, Mammography, Angiography, Atherosclerosis Imaging) covering Medical Image Segmentation, image and volume registration, and physics of medical imaging modalities like: MRI, CT, Xray, PET and Ultrasound. He also holds several United States Patents. Dr. Suri has been listed in Who’s Who 8 times, is a recipient of President’s Gold medal in 1980 and has received more than 50 scholarly and extra-curricular awards during his career. He is also a Fellow of American Institute of Medical and Biological Engineering (AIMBE) and ABI. Dr. Suri’s major interests are: Computer Vision, Graphics and Image Processing (CVGIP), Object Oriented Programming, Image Guided Surgery and Teleimaging. Dr. Suri had worked for Philips Medical Systems and Siemens Medical Research Divisions. He is also a Visiting Professor with Department of Computer Science, University of Exeter, Exeter, England, University of Barcelona, Spain. Currently, Dr. Suri is Senior Director, Research and Development, Fischer Imaging Corporation.

xiv

Professor Chun Yuan is a researcher in the field of magnetic resonance imaging of cardiovascular systems. He has pioneered multiple highresolution MRI techniques for imaging vulnerable atherosclerotic plaques, directed numerous studies examining carotid atherosclerosis with MRI, and has published extensively on the development of imaging sequences and processing techniques and new conceptual quantitative tools for biomedical use. He is a founding member of the Society for Cardiovascular Magnetic Resonance and sits on the Advisory Boards of the Vulnerable Plaque Organization, Pfizer Atherosclerosis, and the Center for Transnational Media Studies. He is presently Professor of Radiology in the School of Medicine at the University of Washington, Adjunct Professor of Electrical Engineering and Bioengineering in the College of Engineering at the University of Washington, the Founder and Executive Director of the Vascular Imaging Laboratory at the University of Washington Medical Center, Visiting Professor at the Post Graduate Medical College of the Chinese Military Service, and has been a Visiting Professor at the University of Lyon and Baylor Medical College. He earned his Ph.D. in Medical Biophysics and Computing at the University of Utah – Salt Lake City. Currently Dr. Yuan designing new technological imaging processes to quantify the morphological and compositional aspects of atherosclerosis. He is also working on making the Vascular Imaging Lab a model interdisciplinary hub that brings together scholars from a broad range of disciplines – from Engineering to Anthropology, from Medicine to Philosophy – to study both the medical and social dimensions of cardiovascular disease. Professor David Wilson is a Professor of Biomedical Engineering and Radiology, Case Western Reserve University. He has research interests in image analysis, quantitative image quality, and molecular imaging, and he has a significant track record of federal research funding in these areas. He has over 60 refereed journal publications and has served as a reviewer for several leading journals. Professor Wilson has six patents and two pending patents in medical imaging. Professor Wilson has been active in the development of international conferences; he was Track Chair at the 2002 EMBS/BMES conference, and he was Technical Program Co-Chair for the 2004 IEEE International Symposium on Biomedical Imaging. Professor Wilson teaches courses in biomedical imaging, and biomedical image processing and analysis. He has advised many graduate and undergraduate students, all of whom are quite exceptional, and has been primary research advisor for over 16 graduate students since starting his academic career. Prior to joining CWRU, he worked in X-ray imaging at Siemens

xv

Medical Systems at sites in New Jersey and Germany. He obtained his PhD from Rice University. Professor Wilson has actively developed biomedical imaging at CWRU. He has led a faculty recruitment effort, and he has served as PI or has been an active leader on multiple research and equipment developmental awards to CWRU, including an NIH planning grant award for an In Vivo Cellular and Molecular Imaging Center and an Ohio Wright Center of Innovation award. He can be reached at [email protected]. Professor Swamy Laxminarayan currently serves as the Chief of Biomedical Information Engineering at the Idaho State University. Previous to this, he held several senior positions both in industry and academia. These have included serving as the Chief Information Officer at the National Louis University, Director of the Pharmaceutical and Health Care Information Services at NextGen Internet (the premier Internet organization that spun off from the NSF sponsored John von Neuman National Supercomputer Center in Princeton), Program Director of Biomedical Engineering and Research Computing and Program Director of Computational Biology at the University of Medicine and Dentistry in New Jersey, Vice-Chair of Advanced Medical Imaging Center, Director of Clinical Computing at the Montefiore Hospital and Medical Center and the Albert Einstein College of Medicine in New York, Director of the VocalTec High Tech Corporate University in New Jersey, and the Director of the Bay Networks Authorized Center in Princeton. He has also served as an Adjunct Professor of Biomedical Engineering at the New Jersey Institute of Technology, a Clinical Associate Professor of Health Informatics, Visiting Professor at the University of Brno in Czech Republic and an Honorary Professor of Health Sciences at Tsinghua University in China. As an educator, researcher and technologist, Prof. Laxminarayan has been involved in biomedical engineering and information technology applications in medicine and health care for over 25 years and has published over 250 scientific and technical articles in international journals, books and conferences. His expertise are in the areas of biomedical information technology, high performance computing, digital signals and image processing, bioinformatics and physiological systems analysis. He is the co-author of the book on State-of-the-Art PDE and Level Sets Algorithmic Approaches to Static and Motion Imagery Segmentation published by Kluwer Publications and the book on Angiography Imaging: State-of-the-Art Acquisition, Image Processing and Applications Using Magnetic Resonance, Computer Tomography, Ultrasound and X-ray, Emerging Mobile E-Health Systems, published by the CRC Press and two volumes of the Handbook of Biomedical Imaging to be published by the Kluwer publications. He has also authored as the editor/co-editor of 20 international conferences and has served as a keynote speaker in international conferences in 43 countries. He is the Founding Editor-in-Chief and Editor Emeritus of the IEEE Transactions on Information Technology in Biomedicine. He served as an elected member of the administrative and executive committees in the IEEE Engineering in Medicine and Biol-

xvi

ogy Society and as the Society’s Vice President for 2 years. His other IEEE roles include his appointments as Program Chair and General Conference Chair of about 20 EMBS and other IEEE Conferences, an elected member of the IEEE Publications and Products Board, member of the IEEE Strategic Planning and Transnational Committees, Member of the IEEE Distinguished Lecture Series, Delegate to the IEEE USA Committee on Communications and Information Policy (CCIP), U.S. Delegate to the European Society for Engineering in Medicine, U.S. Delegate to the General Assembly of the IFMBE, IEEE Delegate to the Public Policy Commission and the Council of Societies of the AIMBE, Fellow of the AIMBE, Senior Member of IEEE, Life Member, Romanian Society of Clinical Engineering and Computing, Life Member, Biomedical Engineering Society of India, U.S. Delegate to IFAC and IMEKO Councils in TC13. He was recently elected to the Administrative Board of the International Federation for Medical and Biological Engineering, a worldwide organization comprising 48 national members, overseeing global biomedical engineering activities. He was also elected to serve as the Publications Co-Chairman of the Federation. His contributions to the discipline have earned him numerous national and international awards. He is a Fellow of the American Institute of Medical and Biological Engineering, a recipient of the IEEE 3rd Millennium Medal and a recipient of the Purkynje award from the Czech Academy of Medical Societies, a recipient of the Career Achievement Award, numerous outstanding accomplishment awards and twice recipient of the IEEE EMBS distinguished service award. He can be reached at [email protected].

xvii

Acknowledgements

This book is the result of collective endeavours from several noted engineering and computer scientists, mathematicans, medical doctors, physicists, and radiologists. The editors are indebted to all of their efforts and outstanding scientific contributions. The editors are particularly grateful to Drs. Petia Reveda, Chun Yuan, David Wilson, Chris L. de Korte, Horst Bischof, Richard Mongrain, John A. Rumberger, Peder C. Pedersen, Costas Pattichi, Anuja Nair, Milan Sonka, Andreas Wahle, William S. Kerwin, Shaoxiong Zhang, and Peter Yim and their team members for working with us so closely in meeting all of the deadlines of the book. We would like to express our appreciation to IOS Press, The Netherlands for helping create this invitational book in plaque imaging. We are particularly thankful to Einar Fredriksson, Director, publications division of IOS press, Anne Marie de Rover, book production coordinator, and Carry Koolbergen editorial administration, rights and permissions, The Netherlands for their excellent coordination of the book at every stage. Dr. Suri would like to thank Fischer Imaging Corporation for their encouragements during his experiments and research. Special thanks are due to Harris Ravine, Roman Janer, and Janine Broda. Thanks go to Philips Medical Systems for their data sets during their experiments. Thanks also go to: Dr. Larry Kasuboski and Dr. Elaine Keeler from Philips Medical Systems, Inc., for their support and motivations. Thanks are also due to my past Ph.D. committee research professors, particularly Professors Linda Shapiro, Robert M. Haralick, Dean Lytle and Arun Somani, for their encouragements. We extend our appreciations to Drs. Ajit Singh, Siemens Medical Systems, George Thoma, Chief Imaging Science Division from National Institutes of Health, Dr. Sameer Singh, University of Exeter, UK for his motivations. Special thanks go to the Book Series Editor, Professor Evangelia Micheli-Tzanakou for advising us on all aspects of the book. We thank the IEEE Press, Academic Press, Springer Verlag Publishers, and several medical and engineering journals for permitting us to use some of the images previously published in these journals. Finally, Jasjit Suri would like to thank his wife Malvika Suri for all the love and support she has showed over the years and to our baby Harman whose presence is always a constant source of pride and joy. I also express my gratitude to my father, a mathematician, who inspired me throughout my life and career, and to my late mother, who most unfortunately passed away a few days before my Ph.D. graduation, and who so much wanted to see me write this book. Special thanks to Pom Chadha and his family, who taught me life is not just books. He is one of my best friends. I would like to also thank my in-laws who have a special place for me in their hearts and have shown lots of love and care for me. David Wilson would like to acknowledge the support of the Department of Biomedical Engineering, Case Western Reserve University in this endeavor. Special thanks are

xviii

due to the many colleagues and students who make research in biomedical engineering an exciting, wondrous endeavor. Swamy Laxminarayan would like to express his loving acknowledgements to his wife Marijke and to his kids, Malini and Vinod for always giving the strength of mind amidst all life frustrations. The book kindles fondest memories of my late parents who made many personal sacrifices that helped shape our careers and the support of my family members who were always there for me when I needed them most. I have shared many ideas and thoughts on the book with numerous of my friends and colleagues in the discipline. I acknowledge their friendship, feedbacks and discussions with particular thanks to Prof. David Kristol of the New Jersey Institute of Technology, Peter Brett of Ashton University, Ewart Carson of the City University, London, Laura Roa of the University of Sevilla in Spain, and Jean Louis Coatrieux of the University of Rennes in France, for their constant support over the past two decades. Chun Yuan would like to acknowledge the support of my colleagues at the Vascular Imaging Lab and the University of Washington – none of this would be possible without their dedication, hard-work, and patience. I thank the National Heart and Lung Institute, Pfizer, and Astra Zeneca for the critical financial support they have provided. And finally, I would also like to express my gratitude to my friends and family for their emotional sustenance.

xix

The Contributors Jasjit S. Suri, Ph.D. Case Western Reserve University Cleveland, OH, USA

Stavros Kakkos, Ph.D. Imperial College University of London London, UK

Chun Yuan, Ph.D. University of Washington Seattle, WA, USA

Maura Griffin, Ph.D. Imperial College University of London London, UK

David L. Wilson, Ph.D. Case Western Reserve University Cleveland, OH, USA

Andrew Nicolaides, Ph.D. Imperial College University of London London, UK

Swamy Laxminarayan, DSc. Idaho State University Pocatello, ID, USA

Jaume Amores, Ph.D. Centre de Visió per Computador Barcelona, Spain

Thomas S. Hatsukami, M.D. University of Washington Seattle, WA, USA Jianming Cai, Ph.D. University of Washington Seattle, WA, USA Efthyvoulos Kryiacou, Ph.D. University of Cyprus Nicosia, Cyprus

Petia Radeva, Ph.D. Centre de Visió per Computador Barcelona, Spain Radj A. Baldewsing, M.Sc Thoraxcenter, Erasmus MC Rotterdam, The Netherlands Johannes A. Schaar, M.D. Thoraxcenter, Erasmus MC Rotterdam, The Netherlands

Marios S. Pattichis, Ph.D. University of New Mexico Albuquerque NM, USA

Chris L. de Korte, Ph.D. Laboratory of Clinical Physics Institution, UMC St. Radboud Nijmegen, The Netherlands

Christodoulos I. Christodoulou, Ph.D. University of Cyprus Nicosia, Cyprus

Frits Mastik Thoraxcenter, Erasmus MC Rotterdam, The Netherlands

Constantinos S. Pattischis, Ph.D. University of Cyprus Nicosia, Cyprus

Patrick W. Serruys, Ph.D. Thoraxcenter, Erasmus MC Rotterdam, The Netherlands

xx

Antonius F. W. van der Steen, Ph.D. Thoraxcenter, Erasmus MC Rotterdam, The Netherlands

Dongxiang Xu, Ph.D. University of Washington Seattle, WA, USA

Jerold S. Shinbane, M.D. UCLA CA, USA

William S. Kerwin, Ph.D. University of Washington Seattle, WA, USA

Matthew J. Budoff, M.D. UCLA CA, USA

Shaoxing Zhang, Ph.D./M.D. CWRU OH, USA

John A. Rumberger, M.D. The Ohio State University Columbus, OH, USA

Olivier Salvado, M.S. CWRU OH, USA

Peder C. Pedersen, Ph.D. Worcester Polytechnic Institute Worcester, MA, USA

Yiping Chen, Ph.D. CWRU OH, USA

Ruben Lara-Montalvo, Ph.D. Worcester Polytechnic Institute Worcester, MA, USA

Claudio Hillenbrand, Ph.D. CWRU OH, USA

Jacob Chakareski, Ph.D. Worcester Polytechnic Institute Worcester, MA, USA

Frank K. Wacker, Ph.D. CWRU OH, USA

Anuja Nair, Ph.D. Cleveland Clinic Foundation Cleveland, OH, USA

Jeffrey L. Durek, Ph.D. CWRU OH, USA

John D. Klingensmith, Ph.D. Cleveland Clinic Foundation Cleveland, OH, USA

Jonathan S. Lewin, M.D. CWRU OH, USA

D. Geoffrey Vince, Ph.D. Cleveland Clinic Foundation Cleveland, OH, USA

Baowei Fei, Ph.D. CWRU OH, USA

Andreas Wahle, Ph.D. University of Iowa Iowa City, IA, USA

Horst Bischof, Ph.D. Graz University of Technology Austria

Milan Sonka, Ph.D. University of Iowa Iowa City, IA, USA

Pierre Elbischger, MSc Graz University of Technology Austria

xxi

Gerald Holzapfel, Ph.D. Graz University of Technology Austria

J. Kevin J. DeMarco, M.D. UMDNJ New Brunswick, NJ, USA

Peter Regitnig, M.D. Graz Medical University Austria

Juan R. Cebral, Ph.D. UMDNJ New Brunswick, NJ, USA

Rosaire Mongrain, Ph.D. McGill University Montreal, Quebec, Canada Richard Leask, Ph.D. McGill University Montreal, Quebec, Canada Ramses Galaz, MEng McGill University Montreal, Quebec, Canada Adrian Ranga, BEng. McGill University Montreal, Quebec, Canada Jean Brunette, Ph.D. Université de Montréal Montreal, Quebec, Canada Anil Joshi, MSc. University of Toronto Toronto, Ontario, Canada Jean-Claude Tardif, M.D. Université de Montréal Montreal, Quebec, Canada

Marcelo A. Castro, Ph.D. George Mason University Farifax, VA, USA Isam Faik, MEng McGill University Quebec, Canada Neil Bulman-Fleming, MEng McGill University Quebec, Canada Trongtin Nguyen, MEng McGill University Quebec, Canada Chunming Li, Ph.D. University of Connecticut Storrs, CT, USA Jim Macione, Ph.D. University of Connecticut Storrs, CT, USA

Olivier F. Bertrand, M.D., Ph.D. Laval University Quebec City, Quebec Canada

Zhi Yang, Ph.D. University of Connecticut Storrs, CT, USA

Peter Yim, Ph.D. UMDNJ New Brunswick, NJ, USA

Martin D. Fox, Ph.D., M.D. University of Connecticut Storrs, CT, USA


xxiii

Contents Preface

vii

The Editors

xiii

Acknowledgements

xvii

The Contributors

xix

Plaque Imaging Using Ultrasound, Magnetic Resonance and Computer Tomography: A Review Jasjit S. Suri, Constantinos S. Pattichis, Chunming Li, Jim Macione, Zhi Yang, Martin D. Fox, Dee Wu and Swamy Laxminarayan Medical Image Retrieval Based on Plaque Appearance and Image Registration Jaume Amores and Petia Radeva MRI Plaque Tissue Characterization and Assessment of Plaque Stability Chun Yuan, Thomas S. Hatsukami and Jianming Cai Intravascular Ultrasound Elastography: A Clinician’s Tool for Assessing Vulnerability and Material Composition of Plaques Radj A. Baldewsing, Johannes A. Schaar, Chris L. de Korte, Frits Mastik, Patrick W. Serruys and Antonius F.W. van der Steen Computer Vision Analysis of Collagen Fiber Bundles in the Adventitia of Human Blood Vessels Pierre J. Elbischger, Horst Bischof, Gerhard A. Holzapfel and Peter Regitnig

1

26

55

75

97

Image Based Biomechanics of Coronary Plaque Rosaire Mongrain, Richard L. Leask, Ramses Galaz, Adrian Ranga, Jean Brunette, Anil Joshi, Jean-Claude Tardif and Olivier F. Bertrand

130

Computed Tomographic Cardiovascular Imaging Jerold S. Shinbane and Matthew J. Budoff

148

Tomographic Plaque Imaging with CT John Rumberger

182

Absolute Measurement of Integrated Backscatter from Arterial Wall Structures Peder Pedersen, Ruben Lara-Montalvo and Jacob Chakareski

208

xxiv

Ultrasound Imaging in the Analysis of Carotid Plaque Morphology for the Assessment of Stroke Efthyvoulos Kyriacou, Marios S. Pattichis, Christodoulos I. Christodoulou, Constantinos S. Pattichis, Stavros Kakkos, Maura Griffin and Andrew Nicolaides

241

On the Assessment of Texture Feature Descriptors in Intravascular Ultrasound Images: A Boosting Approach to a Feasible Plaque Classification Oriol Pujol and Petia Radeva

276

Real-Time Plaque Characterization and Visualization with Spectral Analysis of Intravascular Ultrasound Data Anuja Nair, Jon D. Klingensmith and D. Geoffrey Vince

300

Coronary Plaque Analysis by Multimodality Fusion Andreas Wahle and Milan Sonka

321

Imaging of Plaque Cellular Activity with Contrast Enhanced MRI William Kerwin

360

Inter- and Intra-Observer Variability Assessment of in Vivo Carotid Plaque Burden Quantification Using Multi-Contrast Dark Blood MR Images Shaoxiong Zhang, Jasjit S. Suri, Olivier Salvado, Yiping Chen, Frank K. Wacker, David L. Wilson, Jeffrey L. Duerk and Jonathan S. Lewin Three-Dimensional Volume Registration of Carotid MR Images Baowei Fei, Jasjit S. Suri and David L. Wilson Characterization of Shear Stress on the Wall of the Carotid Artery Using Magnetic Resonance Imaging and Computational Fluid Dynamics Peter Yim, Kevin DeMarco, Marcelo A. Castro and Juan Cebral

384

394

412

Numerical Modeling of Coronary Drug Eluting Stents Rosaire Mongrain, Richard Leask, Jean Brunette, Iam Faik, Neil Bulman-Feleming and T. Nguyen

443

Author Index

461


1

Plaque Imaging Using Ultrasound, Magnetic Resonance and Computer Tomography: A Review Jasjit S. SURI a , Constantinos S. PATTICHIS b , Chunming LI c , Jim MACIONE c , Zhi YANG c , Martin D. FOX c , Dee WU d and Swamy LAXMINARAYAN e a Fischer Imaging Corporation, Denver, CO, USA b Department of Computer Science, University of Cyprus, Cyprus c University of Connecticut, Storrs, USA d Department of Radiological Sciences, University of Oklahoma Health Sciences Center, OK, USA e Idaho State University, Pocatello, ID, USA Abstract. Different classifications have been proposed in the literature for the characterization of atherosclerotic plaque morphology, resulting in considerable confusion. For example plaques containing medium of high level uniform echoes were classified as homogeneous by others and correspond closely to dense and calcified plaques, other types. This survey is to understand different types of plaque when imaged using ultrasound and MR. Keywords. Carotid artery, magnetic resonance angiography, phase-contrast magnetic resonance imaging, computational fluid dynamics, shear stress

1. Ultrasound Vascular Imaging Ultrasound is widely used in vascular imaging because of its ability to visualise body tissue and vessels in a non invasive and harmless way and to visualise in real time the arterial lumen and wall, something that is not possible with any other imaging technique. B-mode ultrasound imaging can be used in order to visualise arteries repeatedly from the same subject in order to monitor the development of atherosclerosis. Monitoring of the arterial characteristics like the vessel lumen diameter, the intima media thickness (IMT) (see Fig. 1) of the near and far wall and the morphology of atherosclerotic plaque (see Fig. 2) are important in order to assess the severity of atherosclerosis and evaluate its progression [1]. The arterial wall changes that can be easily detected with ultrasound are the end result of all risk factors (exogenous, endogenous and genetic) known and unknown and are better predictors of risk than any combination of conventional risk factors. Extracranial atherosclerotic disease, known also as atherosclerotic disease of the carotid bifurcation has two main clinical manifestations a) asymptomatic bruits and b) cerebrovascular

2

J.S. Suri et al. / Plaque Imaging: A Review

a) No filtering

b) After despeckling

Figure 1. B-mode ultrasound imaging of the carotid artery illustrating the IMT measurement before and after despeckle filtering. a) No filtering: IMTaver = 0.837mm, IMTmax = 1.024mm, IMTmin = 0.663mm, IMTmedian = 0.783mm. b) After despeckling: IMTaver = 0.845mm, IMTmax = 1.030mm, IMTmin = 0.667mm, IMTmedian = 0.788mm.

a) Snakes segmentation

b) Expert’s segmentation

Figure 2. Ultrasound imaging segmentation of atherosclerotic carotid plaque at the far wall.

syndromes such as amaurosis fugax, transient ischaemic attacks (TIA) or stroke which are often the result of plaque erosion or rupture with subsequent thrombosis producing occlusion or embolisation [2,3]. Carotid plaque is defined as a localized thickening involving the intima and media in the bulb, internal carotid, external carotid or common femoral arteries (Fig. 2). Recent studies involving angiography, high resolution ultrasound, thrombolytic therapy, plaque pathology, coagulation studies and more recently molecular biology have implicated atherosclerotic plaque rapture as a key mechanism responsible for the development of cerebrovascular events [4–6]. Atherosclerotic plaque rapture is strongly related to the morphology of the plaque [7]. The development and continuing technical improvement of non invasive high resolution vascular ultrasound enables the study of the presence, rate of progression or regression of plaques and most importantly their consistency. The ultrasonic characteristics of unstable (vulnerable) plaques have been determined [8,9] and populations or individuals at increased risk for cardiovascular events can now be identified [10]. In addition, high resolution ultrasound enables the identification of the different ultrasonic characteristics of unstable carotid plaques associated with amaurosis fugax, TIAs, stroke and different patterns of CT-brain infraction [8,9]. This information has provided new insight into the pathophysiology of the different clinical manifestations of extracranial atherosclerotic cerebrovascular disease using noninvasive methods.


3

2. Image Analysis Visual assessment of vascular images and/or video on the monitor of the ultrasound machine is widely used in clinical practice. In recent years, digital image analysis techniques facilitate the possibility of extracting additional useful information in quantitative form, enabling the early diagnosis, disease monitoring, and better treatment. In this section, a brief overview of image processing and analysis techniques like despeckle filtering, normalization, segmentation, feature extraction and selection, and classification are presented. 2.1. Pre-Processing: Normalization In order to extract comparable results when processing images obtained by different operators and equipment and vascular imaging laboratories a standardization method was proposed using blood and adventitia as reference points [11]. The images are standardized manually by adjusting the image so that the median gray level value of the blood is 0–5, and the median gray level value of the adventitia (artery wall) is 180–190 [11]. The image is then linearly adjusted between the two reference points, blood and adventitia. This simple procedure facilitates standardized quantitative analysis for vascular imaging feature extraction and classification. 2.2. Preprocessing: Despeckle Filtering Ultrasound images show a granular appearance known as speckle, which is a form of locally correlated multiplicative noise reducing image contrast and detailed resolution, corrupting medical ultrasound imaging making visual observation difficult, and the signal difficult to detect [12]. Different speckle techniques have been introduced in the literature that are based on local statistics [13], linear scaling of the gray level values [14], the most homogeneous neighbourhood around each pixel [15], geometric filtering [16], homomorphic filtering [17], anisotropic and speckle anisotropic diffusion [18], coherence enhancing diffusion [19] and wavelet filtering [20]. In two recent comparative studies of despeckle filtering techniques evaluated in a large number of asymptomatic and symptomatic ultrasound images of the carotid artery [21,22], it was shown that the local statistics filter is more suitable for the analysis of plaque morphology and texture analysis, whereas the homogeneous mask area filter is more suitable for measuring the intima-media thickness (see Fig. 1) as well as for identifying the degree of stenosis, and the outline of the plaque contour. 2.3. Intima Media Thickness and Plaque Segmentation in Ultrasound Imaging Segmentation in vascular imaging is one of the most difficult tasks in image processing. It targets to subdivide an image into its constituent regions or objects. For example, in the automated segmentation of an ultrasound image of the carotid artery, interest lies in identifying the intima media and subsequently measure its thickness, and furthermore, determine the presence or absence of a plaque, and if there is a plaque determine its contour.

4


For lumen delineation in transversal ultrasound imaging, the Hough transform was investigated [23] as well as to find an initial approximation of the lumen area in the left ventricle [24]. Dynamic programming [25] and cost function optimisation [26] were applied for determining the optimal vessel wall. In intrvascular ultrasound imaging of the carotid artery for detecting the vessel wall the following methods were developed: texture based [27], morphology operators [28], optimal graph searching [29], and dynamic contour-modeling [30]. Furthermore, snakes or deformable models to detect the IMT in 2D [31] and 3D [32] ultrasound images of the carotid artery were developed. These methods are based on the active contour model first introduced by Kass [33] where an active contour is expressed as an energy minimization process, based on internal energy derived from the physical characteristics of the snake based on two components, the continuity energy, and the curvature energy. In general, the snake-based methods require that the initial contour must be drawn by an experienced ultrasonographer. A new method using image normalization as described above, despeckle filtering, automatic initial contour estimation, and snakes for the measurement of IMT was proposed [34] The method requires minimum user interaction and was evaluated on 100 longitudinal ultrasound images with measurements carried out by two experts. The results can be summarized as follows: i) there is no significant difference for the measurement of IMT between the manual and the snakes segmentation measurements, and ii) better segmentation results (smaller inter-observer variability, and smaller coefficient of variation) were obtained for the normalized despeckled images. The manual measurements were smaller than the automated and this finding was also reported in other studies [25,27,30,32]. The no significant difference between the manual and the automated method, shows that the IMT, an important predictor for myocardial infraction and stroke can be reliably computed automatically. To the best of our knowledge there is no technique published enabling the ultrasound automated segmentation of atherosclerotic carotid plaque. A new method was proposed by our group based on snakes for the segmentation of atherosclerotic carotid plaque [35] that is illustrated in Fig. 2. In this method, the initial estimate of the contour of the plaque is automatically estimated by cross correlating the B-mode image with the blood flow image. This initial contour is then deformed on the B-mode image using snakes to find the final contour as shown in Fig. 2a (that could be visually compared to the expert’s segmentation shown in Fig. 2b). The method was compared with the expert’s segmentation results on 35 images where the true positive fraction, TPF, true negative fraction, TNF, false negative fraction, FNF and false positive fraction, FPF, were 86.4%, 84.0%, 8.5%, and 7% respectively. Furthermore, the similarity kappa index, KI, and the overlap index between the expert’s segmentation results and the snakes segmentation results were 85% and 74% respectively, which are considered very satisfactory. These results are comparable to the manual delineation procedure without requiring manual correction in most of the cases. The limitations of this approach i.e. using the blood flow image to locate the blood borders are the following: i) the blood sometimes hides areas of the tissue (verbarations), and ii) the colour does not always fill up the places where blood has a low speed.


5

Asymptomatic SGLDM(mean) Features

SF Features

Angular second moment Contrast

.00064

Mean

44.0

485.18

29.9

Correlation Variance Homogeneity Entropy

0.78 1078.5 0.21 7.86

Standard Deviation Median Skewness Kurtosis

38.5 1.0 1.8

Symptomatic SGLDM(mean) Features Angular second moment Contrast Correlation Variance Homogeneity Entropy

SF Features 0.0027

Mean

25.8

528.56

Standard Deviation Median Skewness Kurtosis

19.2

0.57 595.9 0.32 6.87

21.9 1.40 2.4

Figure 3. SF and SGLDM(mean) texture features for an asymptomatic and a symptomatic ROI images. In each case, top and bottom images represent near wall and far wall ROI plaques respectively (see also [43]).

2.4. Ultrasound Plaque Feature Extraction and Classification Following the segmentation, texture and morphological features are extracted from the segmented plaque images in order to be used for the characterization of the carotid plaques. These features are usually computed on a region of interest (ROI), for example the region prescribed by the plaque contour that is automatically or manually drawn as shown in Fig. 2. ROI images for an asymptomatic and a symptomatic case are shown in Fig. 3. Some of the most common texture feature algorithms that have been used for ultrasound texture analysis are briefly described. Simple statistical descriptors, SD, are computed that include the ROI mean, median, standard deviation, skewness, and kurtosis values (see Fig. 3, and 4a). The spatial gray level dependence matrices, SGLDM, texture features as proposed by Haralick et al. [36] are the most frequently used texture features. These features are the following: angular second moment, contrast, correlation, inverse difference moment, sum average, variance (sum and difference), and entropy (sum and difference). For a chosen distance d that is usually one pixel and for angles θ = 0◦ , 45◦ , 90◦ and 135◦ four values for each of the above texture measures are computed. The mean (see Fig. 3) and range of these four values are usually computed for each feature, and they are used as two different feature sets. The Gray Level Difference Statistics, GLDS, algorithm [37] uses first order statistics of local property values based on absolute differences between pairs of gray lev-

6


Figure 4. Boxplots of two texture features for class 1) symptomatic and class 2) asymptomatic plaques (see also [43]): a) SF Mean, and b) NGTDM Coarseness. The notched box shows the median, lower and upper quartiles and confidence interval around the median for each feature. The dotted line connects the nearest observations within 1.5 of the inter-quartile range (IQR) of the lower and upper quartiles. Crosses (+) indicate possible outliers with values beyond the ends of the 1.5 × IQR.

els or of average gray levels in order to extract the following texture measures: contrast, angular second moment (see Fig. 4b), entropy, and mean. Amadasun and King [38] proposed the Neighborhood Gray Tone Difference Matrix, NGTDM, in order to extract textural features, which correspond to visual properties of texture. The following features are extracted: coarseness, contrast, busyness, complexity, and strength. The statistical feature matrix SFM [39], measures the statistical properties of pixel pairs at several distances within an image, which are used for statistical analysis. Based on the SFM the following texture features are computed: coarseness, contrast, periodicity, and roughness. For the Laws TEM feature extraction [40,41] vectors of length l = 7, L = (1, 6, 15, 20, 15, 6, 1), E = (−1−4, −5, 0, 5, 4, 1) and S = (−1−2, 1, 4, 1−2−1) are used, where L performs local averaging, E acts as edge detector and S acts as spot detector. If the column vectors of length l are multiplied by row vectors of the same length, we compute the Laws lxl masks. In order to extract texture features from an image, these masks are convoluted with the image and the statistics (e.g. energy) of the resulting image are used to describe texture. Fractal Dimension Texture Analysis, FDTA, is based on the work of Mandelbrot [42] who developed the fractional Brownian motion model in order to describe the roughness of natural surfaces. The Hurst coefficients H (k) [41] are computed for different image resolutions, where a smooth texture-surface is described by a large value of the parameter H whereas the reverse applies for a rough texture-surface. The Fourier Power Spectrum, FPS, computes the radial and angular sum of the sample Fourier power spectrum where coarse texture has high values concentrated near the origin, and in fine texture the values are more spread out. Furthermore plaque imaging morphological analysis was carried out as documented in [43] in this book. Statistical analysis of texture features was carried out for a large number of asymptomatic and symptomatic ultrasound images of carotid atherosclerotic plaques [43–45]. It was shown that asymptomatic plaques tend to be brighter, with less contrast, more homogeneous, smoother, with large areas with small gray tone variations, and more periodical, whereas, in symptomatic plaques texture tends to be darker, with higher contrast, more heterogeneous, more rough and less periodical. These findings are summarized in Table 1. Figure 4 shows boxplots of two texture features and the range of values for the asymptomatic and symptomatic groups. The gray scale mean or median indicates how


7

Table 1. Texture characteristics of asymptomatic vs symptomatic plaques. (From [44], © 2003 IEEE, with permission.) Asymptomatic Plaques

Symptomatic Plaques

Brighter

More dark

Less contrast More smooth More homogeneous More periodical More coarse, i.e. large areas with small gray tone variations

Higher contrast More rough More heterogeneous Less periodical Less coarse, i.e. less local uniformity in intensity

bright (high values) or dark (low values) the image is in average. The sum entropy computed with the SGLDM algorithm, is high when the image intensity in neighbouring pixels is more equal and small when the image intensity is more unequal. Angular Second Moment (ASM) of GLDS is small when the gray level probability density values are very close and large when some values are high and others low. Coarseness computed with the NGTDM algorithm, is high when large areas with small gray tone variations are present in the image and small when there is less local uniformity in density. As illustrated in Fig. 4, asymptomatic plaques tend to be brighter (higher mean or median gray level), have higher sum entropy (i.e. the image intensity in neighbouring pixels is more equal), have lower values for ASM and are more coarse, whereas symptomatic plaques tend to be darker (lower mean or median gray level), have lower sum entropy (i.e. the image intensity in neighbouring pixels is more unequal), have higher values for ASM and are less coarse. In an extensive study carried out by Polak et al. [46] where subjects were followed up for an average of 3.3 years, they found that darker (i.e. hypoechoic) carotid plaques are associated with increased risk of stroke. Also, Elatrozy et al. [47] reported that plaques with gray scale median less than 40 are more related to ipsilateral hemispheric symptoms. Wilhjelm et al. [48] in a study with patients scheduled for endarterectomy, carried out a quantitative comparison between subjective classification of the ultrasound images, first and second order statistical features of the ultrasound imaging plaque, and a histological analysis of the surgically removed plaque. They reported some correlation between the aforementioned three types of data where the feature with the highest discriminatory power was contrast. In automated quantitative methods for classifying vascular imaging patterns, both statistical pattern recognition and artificial neural networks (ANN) were used [43–45]. In statistical pattern recognition the k-nearest-neighbour (KNN) classifier was used, whereas in ANN pattern recognition the unsupervised self-organizing map (SOM) was used [44,45], Probabilistic Neural Networks (PNN), and the Support Vector Machines (SVM) algorithms [43]. A brief description is given for the SOM study [44,45]. Nine different SOM models were developed one for each texture feature set as described above, with the output classified into two classes: asymptomatic because the subject was not connected with ipsilateral hemispheric events or symptomatic because the subject was connected with ipsilateral hemispheric symptoms. The percentage of correct classifications score of the above mentioned texture feature sets using the SOM classifier

8


on the evaluation dataset was computed. The highest diagnostic performance was obtained for the SGLDM range feature set where the percentage of correct classifications was 70%, followed by the NGTDM and TEM feature sets. Furthermore the outputs of the SOM classifiers were combined improving the percentage of correct classifications score to 73%. These results are comparable with the results derived with the KNN, PNN and SVM classifiers [43]. Furthermore, the performance of morphological features for carotid plaque classification was also investigated [43,45]. The findings of these studies showed that the percentage of correct classifications score for certain morphology features was slightly lower than the best texture feature sets.

3. Future Trends 3.1. 3D Vascular Ultrasound Imaging In everyday clinical practise, the ultrasonographer manipulates the transducer and mentally transforms the two dimensional (2D) images into anatomical volume, or structure, or lesion, in order to make a diagnosis. Three dimensional imaging (3D) attempts to provide the ultrasonographer or physician with a more realistic reconstruction and visualization of the 3D structure under investigation. In addition, 3D imaging can provide quantitative measurements of volume, surface distance in vascular anatomy, especially in pathological cases. In vascular imaging, a 3D representation was investigated for the visualization of the carotid artery and the quantification of the atherosclerotic plaque volume and morphology [1,32]. Although 3D vascular imaging is very promising in revealing vascular structure and pathology, more work is needed in the directions of fast and accurate free hand scanning, automated or semi-automated segmentation, real-time and user friendly visualization, and 3D texture analysis [1,32]. 3.2. Ultrasound Plaque Imaging and Genetics Atherosclerosis is a multifactorial disease that makes the process of prevention and disease management highly complex. In addition to the many factors that are useful in assessing an individual’s risk of developing a cardiovascular event, recently, new biochemical markers for cardiovascular disease have been identified such as homocysteine, Creactive protein and fibrinogen. However, further work in this area is needed in order to understand and identify their exact role in disease. High resolution ultrasound imaging offers the potential of determining phenotypes more accurately than using conventional risk factors and clinical events. This is achieved because plaque echodensity can characterize the plaques that are unstable and likely to rupture [49]. The ability to identify this type of plaques and hence high risk individuals also offers the advantage of monitoring plaque stabilization drug therapies and the development of new therapeutic strategies, contributing towards the implementation of the most effective strategy to minimize cardiovascular death, and offering a better service to the citizen.


9

4. Boundary Detection for Ultrasound Carotid Artery Images Using Covariance Matrix of Feature Vectors An edge can be defined as a discontinuity in pixel intensity within an image. Edge detection is one of the most important tasks in image processing and computer vision. In ultrasound images, edges appear on the boundaries of organs, blood vessels, and other tissues. The edge information in ultrasound can be used for 3D reconstruction and also, quantification of organ, lesion or tissue size. Although much research has been done on edge detection [1–8], not much has been done for ultrasound images due to the difficulty caused by speckle noise. Most of the conventional edge detectors, e.g., Canny, LoG, or Sobel operators are based on derivative operations. However, derivative operations are sensitive to noise which limits the utility of the derivative-based edge detectors. In this chapter, we present a new edge detection method, with application to the detection of the boundary of the carotid artery in ultrasound images. Our method avoids taking derivative of images, and therefore is more robust to noise. In our edge detection method, edge operations are not directly performed on the original image. For each pixel, we define a feature vector, which characterizes the pixel more adequately than the pixel value. Thus, we derive a feature vector field from the original image. Then we can apply our edge detection technique [84] to the feature vector field. In ultrasound images, the intensity of an individual pixel may not characterize it adequately. It is necessary to find additional parameters, which can characterize a pixel more precisely or provide more reliable information about the pixel, to be used for edge detection. We call these parameters the features of a pixel, and the vector with them as components is called feature vector. If an appropriate feature vector is chosen, it can be expected that the edge detection result can be improved by applying edge detection operations on the feature vector space. It is well known that the speckle noise in ultrasound images is signal dependent. In ultrasound images, different homogenous regions take on different statistical characteristics. In particular, the pixel value on the two sides of an edge has not only different mean but also different standard deviation. The mean and standard deviation of a pixel is the sample mean and standard deviation of a pixel block centered at the pixel of interest. By using both mean and standard deviation as features, we have more information about the pixels, which makes the pixels on the two sides more distinguishable. 4.1. Statistical Edge Models for Vector Valued Images In this section, we consider edge detection on general vector valued images f(i, j ) for two edge models: step edge and ramp edge models. The results in this section is for vector valued images, and will be used for the feature vector field derived from a gray level image in the next section. 4.2. Step Edge Model For a vector valued image f(i, j ), we consider each vector f(i, j ) as a random vector obeying a certain distribution. For an image I (i, j ) with a step edge, we assume that the random vectors f(i, j ) on the two sides of the edge have different means and the covariance matrices. More specifically, we assume that the random vectors f(i, j ) on the two sides of the edge have the means μ1 and μ2 and the covariance matrices 1 and 2 .

10


The edge in the image f(i, j ) can be characterized by the covariance matrix of a random vector X defined as follows. Let be a circular window centered at the pixel (i, j ). Then X is defined as a vector randomly picked up from the set {f(m, n) : (m, n) ∈ }. If a step edge lies in , then it divides into two parts, 1 and 2 , with N1 and N2 pixels, respectively, as shown in Fig. 1. According to the above assumptions, the vector f(i, j ) is a random vector Xi with mean μi and covariance matrix i when the pixel (i, j ) is in i , for i = 1, 2, respectively. The two sides of the step edge are referred to as homogeneous regions, since the vectors f(i, j ) on each side of the edge follow the same distribution. It can be shown that the random vector X has the mean μX = p1 μ1 + p2 μ2 where p1 =

N1 N1 +N2 , p2

=

(1) N2 N1 +N2 ,

and the covariance matrix

X = E{(X − μ)(X − μ)T } = E{XX T } − μμT = p1 E{X1 X1T } + p2 E{X2 X2T } − (p1 μ1 + p2 μ2 )(p1 μ1 + p2 μ2 )T = p1 E{X1 X1T } − p1 μ1 μT1 + p2 E{X2 X2T } − p2 μ2 μT2 + p1 μ1 μT1 + p2 μ2 μT2 + (p1 μ1 + p2 μ2 )(p1 μ1 + p2 μ2 )T = p1 1 + p2 2 + p1 μ1 μT1 − p12 μ1 μT1 + p2 μ2 μT2 − p22 μ2 μT2 − p1 p2 (μ1 μT2 + μ2 μT1 ) = p1 1 + p2 2 + p1 p2 C(μ1 , μ2 )

(2)

Under the assumptions in Eq. (1) and Eq. (2), the covariance matrix of X is approximated by X ≈ p1 p2 (μ1 − μ2 )(μ1 − μ2 )T p1 p2 C(μ1 , μ2 )

(3)

And therefore, the largest eigenvalue of X , denoted by λmax (X ), can be approximated by λmax (X ) ≈ p1 p2 λmax (C(μ1 , μ2 )) = p1 p2 ||μ1 − μ2 ||2

(4)

Note that p1 + p2 = 1, so we get p1 p2 ||μ1 − μ2 ||2

1 ||μ1 − μ2 ||2 4

(5)

where the equality holds when p1 = p2 = 12 , in which case the center of is on the edge. This leads to an important property that the largest eigenvalue of the covariance matrix X reaches its local maximum when the center of moves across the edge. If 1 = 2 , then the above local maximum property of λmax () still holds without the assumptions in Eq. (1) and Eq. (2). In fact, in this case


11

Figure 5. Step edge model.

= 1 + p1 p2 C(μ1 , μ2 )

(6)

It has been proved in [9] that 1 λmax () ≤ λmax 1 + C(μ1 , μ2 ) 4

(7)

where the equality holds when p1 = p2 = 12 . 4.2.1. Ramp Edge Model In ultrasound images, the edges usually do not appear as step edges. It is more appropriate to model them as ramp edges. Let be a circular window centered at the pixel (i, j ). Similarly as in the step edge case, a random vector X is defined as a vector randomly picked up from the set {f(m, n) : (m, n) ∈ }. If there is a ramp step edge in , the window can be divided into three parts, the two homogeneous regions 1 and 2 as in the step edge case, and a ramp region 3 between them, with N1 , N2 , and N3 pixels, respectively. According to the above assumptions, the vector f(i, j ) is a random vector Xi with mean μi and covariance matrix i when the pixel (i, j ) is in i , for i = 1, 2, 3, respectively. Then it can be shown that the random vector X has the mean μ = p1 μ1 + p2 μ2 + p3 μ3

(8)

and the covariance matrix = p1 1 + p2 2 + p3 3 + +

p1 p2 (μ1 − μ2 )(μ1 − μ2 )T p1 + p2

p3 (μ3 − μ)(μ3 − μ)T p1 + p2

(9)

It has been shown [84] that the largest eigenvalue of the above covariance matrix achieves 3 its maximum when p1 = p2 = 1−p 2 , i.e., when the window is centered at the edge. 4.2.2. Local Covariance Matrix of a Vector Valued Image In the previous subsections, we have shown that the covariance matrix has a property that its largest eigenvalue achieves its maximum when the window is centered at the edge. This property directs us to our edge detection for vector valued images using the local covariance matrix , which will be defined as follows. In practice, the covariance matrix is unknown, but it can be estimated by the local covariance matrix of the sample vectors f(ik , jk ) of the N pixels (ik , jk ), k = 1, . . . , N , in . The local covariance matrix of the feature image f(i, j ) in the window is defined by

12


ˆ =

N T 1 ˆ j ) f(ik , jk ) − μ(i, ˆ j) f(ik , jk ) − μ(i, N

(10)

k=1

ˆ j ) is the local mean of the feature vectors of pixels in , defined by where μ(i, ˆ j) = μ(i,

N 1 f(ik , jk ) N

(11)

k=1

ˆ is an estimate of , λmax () can be estimated by Since the local covariance matrix ˆ It is interesting that, the sample covariance matrix ˆ also has the similar form λmax (). as the covariance matrix given in Eq. (4) ˆ 1 + p2 ˆ 2 + p1 p2 C(μˆ 1 , μˆ 2 ) ˆ i,j = p1

(12)

ˆ 2 are the local covariance matrices, and μˆ 1 and μˆ 2 the sample means of ˆ 1 and where the feature image f(i, j ) on 1 and 2 , respectively. For an ideal case that the feature ˆ1 = ˆ 2 = 0 and ˆ reduce vector f(i, j ) are constant f1 on each side of the edge, then to ˆ i,j = p1 p2 (μˆ 1 , μˆ 2 )

(13)

whose largest eigenvalue λ(i, j ) reaches its local maximum when p1 = p2 = 12 . 4.3. Edge Detection on the Feature Vector Field Since our edge operation is applied to the feature image, instead of the original ultrasound images, it is very important to choose an appropriate feature vector. In this work, we adopt the sample mean and standard deviation as the components of the feature vector. For each pixel (i, j ), we calculate the sample mean and standard deviation of the intensity of the neighboring d × d pixels. The sample mean and standard deviation are denoted by μ(i, ˆ j ) and σˆ (i, j ), respectively, and then define the feature vector f(i, j ) as μ(i, ˆ j) f(i, j ) = (14) σˆ (i, j ) Based on the analysis of the covariance matrix of feature vectors in Section II, our edge detection algorithm is presented and can be described by the the following three steps: 1. Calculate the feature vector for every pixel (i, j ). 2. For each pixel (i, j ), calculate the local covariance matrix of the feature vectors, ˆ i,j , and its maximum eigenvalue λ(i, j ). denoted by 3. Threshold on λ(i, j ) with a threshold T . 4. Find the local maxima of λ(i, j ) for the pixels with λ(i, j ) > T . In this chapter, the local covariance matrix of feature image is calculated on an 5 × 5 pixel block centered at the pixel of interest. We review two types of signal dependent noise models. The first one is multiplicative noise model, for which the observed image can be written as I (i, j ) = s(i, j )η(i, j )

(15)


13

(a)

(b)

(c)

(d)

Figure 6. An edge corrupted by signal dependent noise. (a) The original noisy edge. (b) The largest eigenvalues of the local covariance matrices, with 5 × 5 mask. (c) Magnitude of gradient, with σ = 2.0. (d) Magnitude of gradient, with σ = 3.0.

where η(i, j ) is the noise. For images with multiplicative noise, a standard method to process the image, such as edge detection and noise reduction, is to transform the multiplicative noise to additive noise by taking logarithm of the pixel value. Then the conventional edge detection techniques, most of which are designed for additive noise, can be applied to the transformed image. There is another type of signal dependent noise, which cannot be transformed to additive noise simply by taking logarithm or some other operations. For example, a noise model for ultrasound image was reported in [7] and [8], where the observed image I (i, j ) was modeled as (16) I (i, j ) = s(i, j ) + s(i, j )n(i, j ) where n(i, j ) is the identically independent distributed Gaussian noise. To see the above property of the local covariance matrix, we simulated a noisy edge image and calculated the largest eigenvalue of the local covariance matrix. Figure 6(a) shows a edge corrupted by the noise according to the noise model in Eq. (16). The largest eigenvalue of the local covariance matrix of feature vector, denoted by λ(i, j ), is shown

14


(a)

(b)

(c)

(d)

Figure 7. Comparison of edge detections on an ultrasound carotid artery image. (a) The original ultrasound image of a carotid artery. (b) Edge map obtained by our edge detector. (c) Edge map obtained by Sobel edge detector. (d) Edge map obtained by Canny’s edge detector.

in Fig. 6(b). In this example, a 5 × 5 pixel block is used to compute the local covariance matrices. From Fig. 6(b), we can observe that the largest eigenvalue of the local covariance matrix is much larger at the true edge. For comparison, we plot the magnitude of the gradient of the smoothed image by convolution with Gaussian kernel with standard deviation σ . The magnitude of gradient is plotted in Fig. 6(c) and 6(d), for σ = 2.0 and σ = 3.0, respectively. From Figs 6(b), 6(c), and 6(d), we can see that λ(i, j ), shown in Fig. 6(b), is smoother and has a well defined ridge compared to the magnitude of the smoothed image by convolution, shown in Fig. 6(c) and 6(d). This shows that the edge can be more easily detected by our method by thresholding and non-maxima suppression on λ(i, j ). 4.4. Application to Ultrasound Carotid Artery Images and Performance Evaluation Our algorithm has been applied to ultrasound images of carotid arteries, and compared with the traditional edge detectors, such as Sobel and Canny detectors. For example, Fig. 7(a) shows an ultrasound carotid artery image, and the edge map obtained by our method is shown in Fig. 7(b). For comparison, we also used Sobel and Canny’s edge detectors to get the edge maps of the same image, shown in Fig. 7(c) and 7(d), respectively. By visual comparison of the results, we observed that the edge map obtained by our edge detector had better edge localization, edge continuity, and fewer false edges than the one


15

Figure 8. Performance evaluation of edge detection by Pratt’s figure of merit.

by Sobel and Canny’s detectors. For the image shown in Fig. 7(d), we used a Gaussian kernel with σ = 3.0 for the Canny’s edge detector. In order to quantitatively evaluate our edge detection algorithm, Pratt’s figure of merit (FOM), was used to compare our and Canny’s methods. The FOM is defined as D 1 1 max{ND , NI } 1 + α(di )2

N

FOM =

(17)

i=1

where di is the distance between a declared edge pixel and the nearest true edge pixel, α is a calibration constant, and NI and ND are the ideal and the detected edge pixels respectively. In order to use FOM as a metric to compare the edge detection performances, we used simulated images so that we are able to know the location of the true edge and therefore compute the metric. The noise performance of the edge detectors was evaluated for images with different signal-to-noise ratios (SNR). We use the definition of the SNR of an edge image in [4] as below SNR = 20 log

k σ

(18)

where k is the edge contrast and σ is the standard deviation of the additive Gaussian noise. Figure 10 plots the FOM vs. SNR for our and Canny’s edge detection methods. For Canny’s edge detector, we used the Gaussian kernel with σ = 2.0. The evaluation by FOM shows that our method has better performance, especially for the case of low signal-to-noise ratio. 4.5. Conclusion and Future Work In this chapter, we present a new edge detection method using the covariance matrix of feature vector. The edge detection is performed on the feature image calculated from the original gray level image. Our edge detection algorithm has been implemented and applied to real and simulated images. Pratt’s figure of merit was used to quantitatively evaluate the performance or our and Canny’s edge detectors. The results show that our method is superior to the conventional edge detection methods, especially for the case of low signal-to-noise ratio. In our future work, we will try to find more types of feature vectors for ultrasound images of different organs or tumors using different statistical parameters, according to the distributions of pixel intensity shown by different objects in the images.

16


5. MR Angiography For ease of explanation, we classify MRA into four subtechniques: 1) Time-of-Flight imaging, 2) Phase Contrast Angiography, 3) Contrast Enhanced MRA, and 4) Black Blood Imaging. All four methods have their particular advantages and limitations. All four also benefit from improved RF coil design [90] as described later in the book. Our goal in this section is to describe the advantages and challenges of each technique in the context of clinical evaluation of disease. Of particular note are black blood imaging techniques (Section 5.1.4) which are used to evaluate plaque formations. 5.1. Time-of-Flight (TOF) Imaging TOF Imaging is clinically easy to implement and can produce images without contrast agents. This bright blood technique relies on the inherent sensitivity to MR physics of moving spins because fresh flowing spins into the slice/slab have no previous excitation history and thus will have relatively brighter signal. For stationary spins, previous RF excitation in the slice/slab will decrease their signal and result in more saturated signal (i.e. reduced signal). In a nutshell, at locations where there are a greater number of flowing spins, a brighter resultant signal is observed by the TOF effect. MR parameters play an important role in TOF imaging. Short echo times(TE) and short repetition times(TR) are used to enhance the TOF effect. Additionally, the larger the flip angle (FA), the more signal-to-noise ratio is gained (SNR) and overall suppression of stationary tissue is achieved. The user/technologist should be aware that TOF effect on slower moving spins is also more adversely affected with greater FA; therefore, increasing the FA to a certain degree will make contrast stronger with the sacrifice of less vascular detail. This is due to the fact that as FA is increased, the distance blood will travel prior to saturation is shortened. In these conditions, slower flow will receive more saturation. The effects of improved background suppression and SNR increases must be balanced against sensitivity to a smaller flow regime. In summary, slice/slab thickness, distance traveled, relaxation time, flow velocity, and the duration of TE should be considered for TOF imaging. Saturation bands can also be used in specific protocols to selectively suppress arterial or venous flow. A band can be placed superiorly or inferiorly to the direction of flow. The purpose of this band is to selectively suppress either descending flow (typically venous) with a superior saturation band, or ascending flow (typically arterial) with an inferior saturation band. In Fig. 9, a MR venogram of a 20-year-old post-partum female who is to be screened for venous thrombosis is shown. When saturation pulses are applied (see Fig. 9) less arterial flow is suppressed. This saturation band technique provides potentially better visualization on strictures or stenosis after projection. A limitation of 2D imaging with TOF is the potential creation of mis-registration artifacts. In general, such artifacts are produced when the time of encoding differs between the frequency and phase axes. Figure 10 demonstrates the MR science behind this artifact. In addition, complex flow, turbulence, and the above mis-registration artifact threaten the geometric spatial integrity of vessels with the TOF technique. TOF also offers an advantage as a way to providing angiographic images without the use of a contrast agent, thus saving expense and decreasing inconvenience to the patient, thereby increasing throughput and efficiency of care. Similar to many hospitals,


17

Figure 9. Left: Diagram demonstrating the MR physics diagram of production of bright signal in moving flow using Time-of-Flight Imaging. Note that TR«T1, as the fresh unsaturated spins are bright (red) compared with stationary tissue (blue). Note the acquired slice is dotted around its border. Right 2 panels: Post-partum patient by TOF to evaluate for venous thrombosis. Arrow points to saturated arterial flow when a saturation band is applied (far right).

Figure 10. Left: Blood vessels with flow rate (v) will appear shifted in position. This shifted appearance is known as the mis-registration artifact. Right: Six Slab and Two Slab MOTSA, displaying different levels of slab boundary artifact.

our institution represents a joint partnership between a Physician/Academic group and a managed health care group. Efficient/cost-effective delivery of health care is an essential part of our mission. It is often preferred to acquire these images in a 3D acquisition mode to achieve high-resolution. The advantage of the 3D technique (over a 2D technique) is the capability of high through-plane resolution that is achieved with the boost in SNR. A disadvantage of the 3D technique is the reduced sensitivity to slow flow (due to the thicker slabs) and reduced background suppression. Another major problem of 3D imaging is the slab-boundary-artifact that results from a reduced sensitivity in the application of uniform RF excitation over a wide slab. Limitations in this excitation are due to the fact that there is only a finite time available for short TE pulse sequences. While there have been traditional and sophisticated ways to get around this artifact, the two most notable are the “multiple overlapping thin slab acquisitions” (MOTSA) technique initially popularized by Dennis Parker [88]. Recent developments in TOF slab imaging include “SLiding Interleaved KY” (SLINKY) by Kecheng Liu [87]. 5.2. Contrast Enhanced Magnetic Resonance Angiography (CE-MRA) Another way to perform bright blood imaging is to rely on timing the acquisition with the arrival of the contrast agent bolus during the enhancement period(CE-MRA). The advantages of the CE-MRA technique over TOF are overall shorter scan times, the ability to produce images from different phases of uptake (e.g. arterial and/or venous phase) as shown in Fig. 11, and higher CNR/SNR than conventional TOF. CE-MRA when applied correctly can provide a more accurate depiction of the geometry than conventional TOF.

18


Figure 11. Left: Arterial and venous phase scans using CE-MRA showing to different looks acquired at different times post-contrast injection. Right: Desirable(left) and undesirable(right) timing for bolus chase CE-MRA of carotids.

Figure 12. Moving Table Implementation using CE-MRA for peripheral vascular examination. Produced with assistance of Medical Physics Section at OUHSC.

The major limitation of the CE-MRA techniques is achieving correct bolus timing. Variability of blood flow rates, particularly in different aged individuals with different habitus, causes a challenge in achieving uniform exam conditions and therefore reproducible imaging. Some techniques to improve CE-MRA include the use of elliptic centric ordering, bolus timing (smart prep), and or fluoroscopic timing bolus as described by Martin Prince [89]. Figure 11 illustrates a good timing bolus (left) as compared with a failed bolus timing (right). Achieving desirable timing, is very dependent on the operator gaining confidence in evaluation of the delay in timing and bolus operation. Technical factors should be discussed with basic science faculty and physicians; however, most important is the seasoning of the technologist in regards to selecting proper timing and parameters. When a level of comfort with bolus timing is achieved by the operator, CE-MRA strategies are very useful in run-off studies where the Bolus is chased as the couch bed is moved from station to station. With correct timing one can piece these images together to make full body views for screening. Many vendors now provide the ability to paste together several CE-MRA images together to make the whole-body MRA scans as shown in Fig. 12. These illustrations were provided with assistance of clinical medical physics group at OU Medical Center. 5.3. Phase Contrast Angiography (PCA) PCA is another MRI technique that holds promise in the quantification and evaluation of blood flow. It has utility for a quick and easy localizer, carotid imaging, as well as for quantification of stenotic flow. Another application is for co-arctation of the aorta. An example is for coarctation in which PCA can provide additional information concerning the level of disease by evaluation of the flow rate in the aorta. A pitfall for PCA is the complexity of setting the correct velocity encoding, which provides additional burden to the technologist who may not routinely use PCA. In ad-


19

Figure 13. Velocity and TE play an essential role in the contrast of black blood spin echo images. If distance traveled of flow is faster than TE/2 then can potentially only see the 180 degree pulse = Void signal. For all the beginning arrow is at time TA1, end of arrow is at TA2. Note only case A in the diagram, will receives an adequate 90–180 pair that produces normal signal, all other combinations (B,C,D) lead to signal loss in the moving spins(flow). Also, demonstration of Moya-moya which bright blood imaging and dark blood imaging.

Figure 14. a) Black blood imaging surveys of heart in a coronal b) 4 chamber view of the heart and c) black blood localizer for the thoracic artery showing both the exterior and intimal layer of the vessel.

dition the lengthy time involved in acquiring these images can present a challenge to the use of PCA. The operator should consult with technical staff and receive specialized training (potentially by the MR faculty at our institution). 5.4. Black Blood MRA Black blood images make excellent localizers providing detail inside and outside the vessel. A downside to black blood imaging is its inherent long scan time. A challenge is for projections as minimum intensity projections (MinIP) must be maintained. MinIP techniques in the presence of air structures in lungs and surrounding tissue and surrounding air make it hard to generate fast turn-around with this type of post processing. Figure 13 illustrates black blood imaging in a patient with suspected Moya-moya disease. The increased prevalence in angiogenesis can be easily seen on bright blood images. However, the flow voids are best visualized on black blood images in the areas of faster flow in the insular/border zone region in this patient. Another clinical application for black blood imaging is for thoracic aneurysm (Fig. 13). The thoracic aneurysm on black blood images demonstrates both the interior and exterior wall (not possible to visualize on bright blood imaging). Additionally, misregistration makes using TOF imaging less satisfactory for visualizing the extent of the disease. Black blood imaging has received increasing attention in the last couple of years as popularized by Fayad and Edelman [85,86]. The method relies on the passage of blood flow and the combination of spin echoes. A late TE will provide a greater sensitivity to slower flows than a conventional TOF image. Too great an echo time will of course result in flow loss. The main advantage of black blood is that it benefits from more integrity in the signal voids and does not have the inherent flow misregistration problems that the TOF imaging.

20


Figure 15. Left: an A1 segment aneurysm shown on MIP CTA images. Right: Volume Reconstructed CTA images, illustrating shape and morphology of the pseudoaneurysm.

Figure 16. Post-processed CTA images, specific arterial supplies can be enhanced on projected MIP images.

6. Computed Tomography Angiography (CTA) An alternative to MRA (although a more invasive approach) is to use Computed Tomography Angiography. CTA provides significantly higher resolution and more spatial integrity than MRA techniques. CTA can also achieve higher resolution in matter of seconds while MRA imaging (by TOF/CE-MRA) can require approximately 30–40 seconds. The drawbacks to CTA are in the large X-ray dose, the poorer visibility near bone (beam hardening), and the separation of arterial/venous phases that is achievable with MRI TOF imaging. Nevertheless, due to the 24 hr availability of CTA at many institutions as well as the relatively less invasive nature of the exam (compared to commonly used fluoroscopic DSA techniques) and the ability to perform cross-sectional imaging with the possibility of retrospective reformat of the image. Recent technology enables multiple dimensional reconstruction using C-arms; however, at much radiation dose to the patient. It is the practical goal also to be able to visualize the general morphology of the tumor. Aneurysm present in fusiform or sacular manner, the size of the neck, and the size of aneurysm on a sacular tumor are also important for GDC coiling. Coiling a fusiform aneurysm would lead to obstruction of the entire vessel and inevitable complications. The location of the aneurysm is critical; for example it would be unwise potentially to coil an aneurysm located in a major branch of the MCA. However, these decisions should be discussed as part of the patient and surgeon consultation. With detailed work by the MP and physician visualization team, detailed CTA that can extract definition of vascular arterial phase are shown in Figs 15 and 16. Producing images that aid interpretation is key; at our institution the neurosurgeons receive these images and desire to grasp the 3D nature of the problem quickly. Such 3D visualization is better accepted by the surgeons then the conventional 2D images. CTA images are acquired with so many thousands of slices that to review these images would be exceedingly time consuming without 3D visual summary, though it is always essential to return to conventional 2D raw images for confirmation. Vital to CTA is the ability to have an effective visualization workstation that will handle large numbers of high resolution images. While acquisition is a primary goal, it is important to capture as much of the arterial phase as possible for producing high quality


21

work. Timing is the key in capturing the arterial phase. Retrospective reconstruction that produces large numbers of thin isotropic high resolution images for 3D rendering is needed. The workstation enables 3D MIPs and the ability to cut out venous flow as well as bone. Sufficient familiarity with cerebrovascular anatomy is absolutely a prerequisite. Anatomical fluency and facility with the computer workstation are key components to the success of our team. Extraction of the correct vascular phase can take 25 minutes for a well trained person, but may take longer for the inexperienced. Often the complexity of the anatomy or the time consuming procedures make the evaluation of CTA more difficult in a routine clinical practice, which has relegated the procedure to academic or bigger hospitals that have the appropriate staffing. CTA plays a major role in the analysis of aneurysms. CT angiography also can be used to evaluate vessel integrity and its function. The utilization of CT methods (especially EBCT) which are particularly useful for plaque imaging are discussed in the subsequent chapters of this book.

Acknowledgements The first authors would like to thank all the team members for their contributors to this review chapter. The IVUS contributions were made by Constantinos S. Pattichis, the boundary estimation techniques for IVUS were made by Chunming Li, Jim Macione, Zhi Yang and Martin Fox and MR and CT contributions were made by Dee Wu.

References [1] M. Hennerici, D. Neuerburg-Heusler, Vascular Diagnosis With Ultrasound, Stutgart-New York: Thieme, 1998. [2] M. Sonka, J.M. Fitzpatrick, Handbook of Medical Imaging, Vol. 2, pp. 809–914, Washington USA: SPIE Press, 2000. [3] D.E. Gutstein, V. Fuster, Pathophysiology and clinical significance of atherosclerotic plaque rapture, Cardiovasc. Res., 41:323–333, 1999. [4] A.J. Zukowski, A.N. Nicolaides, R.T. Lewis, A.O. Mansfield, M.A. Williams, E. Helmis, G.M. Malouf, D. Thomas, A. Al-Kutoubi, P. Kyprianou et al., The correlation between carotid plaque ulceration and cerebral infraction seen on CT scan, J. Vasc. Surg., 1:782–786, 1984. [5] P. Libby, Molecular basis of acute coronary syndromes, Circulation, 91:2844–2850, 1995. [6] S. Clinton, R. Underwood, L. Hayes, M.L. Sherman, D.W. Kufe, P. Libby, Macrophage-colony stimulating factor gene expression in vascular cells and human atherosclerosis, Am J Pathol., 140:301–316, 1992. [7] M.J. Davies, P.D. Richardson, N. Woolf, D.R. Katz, J. Mahn, Risk of thrombosis in human atherosclerotic plaques: role of extracellular lipid, macrophage and smooth muscle cell content, Br Heart J, 69:377–38, 1993. [8] A.N. Nicolaides, S. Kakkos, M. Griffin, G. Geroulakos, E. Bashardi, Ultrasound plaque characterisation, genetic markers and risks, Pathophysiol Haemost Thromb, 32(supl 1):1–4, 2002. [9] M.M. Sabetai, T.J. Tegos, A.N. Nicolaides, T.S. El-Atrozy, S. Dhanjil, M. Griffin, G. Belgaro, G. Geroulakos, Hemispheric symptoms and carotid plaque echomorphology, J Vasc Surg, 31:39–49, 2000. [10] T.J. Tegos, M.M. Sabetai, A.N. Nicolaides, T.S. El-Atrozy, S. Dhanjil, J.M. Stevens, Patterns of brain computed tomography infraction and carotid plaque echogenicity, J Vascs Surg, 33:334–339, 2001. [11] G. Belgaro, A.N. Nicolaides, G. Laurora, M.R. Cesarone, M. De Sanctis, L. Incandela, A. Barsoti, Ultrasound morphology classification of the arterial wall and cardiovascular events in a 6-year follow-up study, Atherioscler Thromb Vasc Biol, 16:851–856, 1996.

22


[12] L.M. Reilly, R.J. Lusby, L. Hughes, L.D. Ferell, R.J. Stoney, W.K. Ehrenfeld, Carotid plaque histology using real-time ultrasonography: clinical and therapeutic implications, Am J Surg, 146:188–193, 1983. [13] J.M. Johnson, M.M. Kennelly, D. Decesare, S. Morgan, A. Sparrow, Natural history of asymptomatic carotid plaque, Arch surg, 120:1010–1012, 1985. [14] A.C. Gray-Weale, J.C. Graham, J.R. Burnett, K. Burne, R.J. Lusby, Carotid artery atheroma: comparison of preoperative B-mode ultrasound appearance with carotid endarterectomy specimen pathology, J Cardiovasc Surg, 29:676–681, 1998. [15] B. Widder, K. Paulat, J. Hachspacher, H. Hamann, S. Hutschenreiter, C. Kreutzer, F. Ott, J. Vollmar, Morphological characterisation of carotid artery stenosis by ultrasound duplex scanning, Ultrasound Med Biol, 16:349–354, 1990. [16] J.M. DeBray, J.M. Baud, Dauzat M for the consensus Conference: Consensus on the morphology of carotid plaques, Cerebrovasc Dis, 7:289–296, 1997. [17] J.F. Polak, L. Shemanski, D.H. O’Leary, D. Lefkowitz, T.R. Price, P.J. Savage, W.E Brant, Reid C. for the cardiovascular health study. Hypoechoic plaque at US of the carotid artery: An independent risk factor for incident stroke in adults aged 65 years or older, Radiology, 208:649–654, 1998. [18] N.M. El-Barghouty, A. Nicolaides, V. Bahal, G. Geroulakos, A. Androulakis, The identification of high risk carotid plaque, Eur J Vasc Surg, 11:470–478, 1996. [19] T. Elatrozy, A. Nicolaides, T. Tegos, A. Zarka, M. Griffin, M. Sabetai, The effect of B-mode ultrasonic image standardisation on the echogenicity of symptomatic and asymptomatic carotid bifurcation plaques, International Angiology, Vol. 17, No. 3, pp. 179–186, Sept. 1998. [20] J. W. Goodman. Some fundamental properties of Speckle. Journal of Optical Society of America, vol. 66, no. 11, 1976, pp. 1145–1149. [21] C.B. Burckhardt, Speckle in Ultrasound B-Mode scans, IEEE Trans. on Sonics and Ultrasonics, vol. SU25, no. 1, pp. 1–6, 1978. [22] R.F. Wagner, S.W. Smith, J.M. Sandrik, and H. Lopez, Statistics of Speckle in Ultrasound B-Scans, IEEE Trans. Sonics. Ultrason., vol. 30, pp. 156–163, 1983. [23] R.F. Wagner, M.F. Insana, D.G. Brown, Unified approach to the detection and classification of speckle texture in diagnostic ultrasound, Optical Engineering, vol. 25, no. 6, pp. 738–742, June 1986. [24] R.F. Wagner, M.F. Insana, S.W. Smith, Fundamental Correlation Lengths of Coherent Speckle in Medical Ultrasonic Images, IEEE Trans. On Ultrasonics, Feroelectrics and Frequency Control, vol. 35, no. 1, pp. 34–44, Jan. 1988. [25] J.S. Lee, Speckle analysis and smoothing of synthetic aperture radar images, Computer Graphics and Image Processing, 17: 24–32, 1981. [26] R. Woods, R. Gonzalez, Digital Image Processing, Addison-Wesley, 2002. [27] S.M. Ali, R.E. Burge, New automatic techniques for smoothing and segmenting SAR images, Signal Processing, 14: 335–346, 1988. [28] L. Busse, T.R. Crimmins, J.R. Fienup, A model based approach to improve the performance of the Geometric filtering speckle reduction algorithm, IEEE Ultrasonic Symposium, 1353:1356, 1995. [29] J. Saniie, T. Wang, N. Bilgutay, Analysis of homomorphic processing for ultrasonic grain signal characterization, IEEE Transactions on ultrasonics, ferroelectrics and frequency control, 3:365–375, 1989. [30] Y. Yongjian, S.T. Acton, Speckle Reducing Anisotropic Diffusion, IEEE Transactions on Image Processing, 11:1260–1270, 2002. [31] K. Abd-Elmonien, A.-B. Youssef, Y. Kadah, Real-Time Speckle Reduction and Coherence Enhancement in Ultrasound Imaging via Nonlinear Anisotropic Diffusion, IEEE Trans. Biomed. Eng., vol. 49, no. 9, pp. 997–1014, Sept. 2002. [32] A. Pizurica, W. Phillips, I. Lemahieu, M. Acheroy, Despeckling SAR images using wavelets and a new class of adaptive shrinkage estimators, ICIP 2001, Thessaloniki-Greece, 1:233–236, 2001. [33] C. Loizou, C. Christodoulou, C.S. Pattichis, R. Istepanian, M. Pantziaris, A. Nicolaides, Speckle Reduction in Ultrasound Images of Atherosclerotic Carotid Plaque, Proceedings IEEE 14th International Conference on Digital Signal Processing, 2:525–528, 2002. [34] C. Loizou, C.S. Pattichis, R. Istepanian, M. Pantziaris, E. Kyriakou, T. Tyllis, A. Nicolaides, Ultrasound Image Quality Evaluation, ITAB 2003, Proc of the 4th Annual IEEE Conf on International Technology Applications in Biomedicine, Birmingham UK, 138–141, 2003. [35] S.-M. Wu, Y.-W. Shau, F.-C. Chong, F.-J. Hsieh, Non-invasive assessment of arterial dimension waveforms using gradient-based Hough transform and power Doppler ultrasound imaging, Journal of Medical & Biological Engineering & Computing, 39:627–632, 2001.


23

[36] S. Malassiotis, M. Strintzis, Tracking the left ventricle in echocardiographic images by learning heart dynamics, IEEE Transactions on Medical Imaging, 18:282–290, 1999. [37] I. Wendelhag, Q. Liang, T. Gustavsson, J. Wikstrand, A new automated computerized analyzing system simplifies reading and reduces the variability in ultrasound measurement of intima media thickness, Stroke, 28:2195–2200, 1997. [38] D.S. Meier, R.M. Cothren, D.G. Vince, J.F. Gornhill, Automated morphometry of coronary arteries with digital image analysis of intra-vascular ultrasound, Am. Heart J., 133:681–690, 1997. [39] A. Mojsilovic, M. Popovic, N. Amodaj, R. Babic, M. Ostojic, Automatic segmentation of intravascular ultrasound images: A texture based approach, Anal in Biomed. Eng., 25:1059–1071, 1997. [40] C.J. Bouma, W.J. Niessen, K.J. Zuiderveld, E.J. Gussenhoven, M.A. Viergever, Automated lumen definition from 30 MHz intravascular ultrasound images, Med. Image Anal., 1:363–377, 1997. [41] X. Zhang, C.R. McKay, M. Sonka, Tissue characterization in intra-vascular ultrasound images, IEEE Trans. Med. Imaging, 17:889–899, 1998. [42] F. Mao, J. Gill, D. Downey, A. Fenster, Segmentation of carotid artery in ultrasound images: Method development and evaluation technique, Med. Phys. 27;8:1–10, 2000. [43] Da-chuan Cheng, Arno Schmidt-Trucksaess, Kuo—sheng Cheng, Hans Burkhardt, Using snakes to detect the intimal and adventitial layers of the common carotid artery wall in sonographic images, Computer methods and programs in biomedicine, 67:27–37, 2002. [44] A. Zahalka and A. Fenster, An automated segmentation method for three-dimensional carotid ultrasound images, Physics in Medicine and Biology, 46:1321–1342, 2001. [45] M. Kass, A. Witkin and D. Terzopoulos, Snake: active contour models, Int. Journal Compt. Vision, 1: 321–331, 1998. [46] P.C. Yuen, G.C. Feng and J.P Zhou, A contour detection method: Initialization and contour model, Pattern recognition letters, 20:141–148, 1999. [47] C.P. Loizou, C.S. Pattichis, R.S.H. Istepanian, M. Pantziaris, T. Tyllis, “Intima media segmentation of the carotid artery”, submitted to the IEEE Trans. Information Tech. Biomedicine, 2004. [48] C. Loizou, C. Pattichis, R. Istepanian, M. Pantziaris, A. Nicolaides, Atherosclerotic Carotid Plaque Segmentation, submitted to the 26th Annual Int. Conf. IEEE Eng. in Med. and Biol. (EMBS), Sept. 1–5, 2004, San Francisco, California, USA. [49] R.M. Haralick, K. Shanmugam, I. Dinstein, Texture Features for Image Classificatio, IEEE Transactions on Systems, Man, and Cybernetics, Vol. SMC-3, pp. 610–621, Nov. 1973. [50] J.S. Weszka, C.R. Dyer, A. Rosenfield, A Comparative Study of Texture Measures for Terrain Classification, IEEE Transactions on Systems, Man. and Cybernetics, Vol. SMC-6, pp. 269–285, April 1976. [51] M. Amadasun, R. King, Textural Features Corresponding to Textural Properties, IEEE Transactions on Systems, Man, and Cybernetics, Vol. 19, No 5, pp. 1264–1274, Sept./Oct. 1989. [52] Chung-Ming Wu, Yung-Chang Chen, Statistical Feature Matrix for Texture Analysis, CVGIP: Graphical Models and Image Processing, Vol. 54, No 5, pp. 407–419, Sept. 1992. [53] K.I. Laws, “Rapid Texture Identification”, SPIE, 1980, Vol. 238, pp. 376–380. [54] Chung-Ming Wu, Yung-Chang Chen, Kai-Sheng Hsieh, Texture Features for Classification of Ultrasonic liver Images, IEEE Transactions on Medical Imaging, Vol. 11, No 2, pp. 141–152, June 1992. [55] B.B. Mandelbrot, The Fractal Geometry of Nature, San Francisco, CA, Freeman, 1982. [56] E. Kyriakou, M.S. Pattichis, C. Christodoulou, C.S. Pattichis, S. Kakkos, A. Nikolaides, Ultrasound Imaging in the Analysis of Carotid Plaque Morphology for the Assessment of Stroke, in Advanced Plaque Imaging: Pixel to Molecular Levels, Ed. by J.S. Suri, C. Yuan, D.L. Wilson, S. Laxminarayan, CRC Press, 2004. [57] C. Christodoulou, E. Kyriacou, M.S. Pattichis, C.S., Pattichis, A. Nicolaides, A Comparative Study of Morphological and other Texture Features for the Characterization of Atherosclerotic Carotid Plaques, Computer Analysis of Images and Patterns, LNCS2756, 10th International Conference, CAIP 2003, Ed. by N. Petkov, and M. Westenberg, Springer, Groningen, The Netherlands, August 25–27, pp. 503–511, 2003. [58] P. Maragos, Pattern spectrum and multiscale shape representation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 11:701 715, 1989. [59] C.I. Christodoulou, C.S. Pattichis, M. Pantziaris, A. Nicolaides, Texture Based Classification of Atherosclerotic Carotid Plaques, IEEE Transactions on Medical Imaging, Vol. 22, No 7, pp. 902–912, July 2003. [60] C. Christodoulou, C.S. Pattichis, E. Kyriacou, M.S. Pattichis, M. Pantziaris, A. Nikolaides, Texture and

24

[61]

[62] [63]

[64] [65]

[66]

[67]

[68]

[69]

[70]

[71] [72] [73] [74] [75]

[76] [77] [78] [79] [80] [81]

J.S. Suri et al. / Plaque Imaging: A Review Morphological Analysis of Ultrasound Images of the Carotid Plaque for the Assessment of Stroke, in Applied Medical Image Analysis Methods, Ed. by L. Costaridou, CRC Press, 2004. C.I. Christodoulou, C.S. Pattichis, E. Kyriacou, A. Nicolaides, Selection of Parameters for Texture Analysis for the Classification of Carotid Plaques, submitted to Medicon 2004 Health in the Information Society, X Meditteranean Conference on Medical and Biological Engineering, July 31 – August 5, Naples 2004. T. Elatrozy, Nicolaides A., T. Tegos, M. Griffin, The Objective Characterization of Ultrasonic Carotid Plaque Features, Eur J Vasc Endovasc Surg 16, pp. 223–230, 1998. J.E. Wilhjelm, L.M. Gronholdt, B. Wiebe, S.K. Jespersen, L.K. Hansen, H. Sillesen, “Quantitative Analysis of Ultrasound B-Mode Images of Carotid Atherosclerotic Plaque: Correlation with Visual Classification and Histological Examination”, IEEE Transactions on Medical Imaging, Vol. 17, No. 6, pp. 910–922, December 1998. J. Kittler, M. Hatef, R. Duin, J. Matas, On Combining Classifiers, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 20, No 3, pp. 226–239, March 1998. A. Fenster and D.B. Downey. Three-Dimensional Ultrasound Imaging, in J. Beutel, H.L. Kundel, R.L. Van Metter (eds), Handbook of Medical Imaging: Vol. 1. Physics and Psychophysics. Washington: SPIE Press, 2000, pp. 463–509. T.H. Karson, R.C. Zepp, S. Chandra, A. Morehead, and J.D. Thomas, Digital Storage of Echocardiograms Offers Superior Image Quality To Analog Storage, Even with 20:1 Digital Compression: Results of the Digital Echo Record Access Study, Journal of the American Society of Echocardiography, vol. 9, pp. 769–778 (1996). J.S. Soble, G. Yurow, R. Brar, T. Stamos, A. Neumann, M. Garcia, M.F. Stoddard, P.K. Cherian, B. Bhamb, and J.D. Thomas, Comparison of MPEG Digital Video with Super VHS Tape for Diagnostic Echocardiographic Readings, Journal of the American Society of Echocardiography, vol. 11, pp. 819–825 (1998). M.L. Main, D. Foltz, M.S. Firstenberg, E. Bobinsky, D. Bailey, B. Frantz, D. Pleva, M. Baldizzi, D.P. Meyers, K. Jones, M.C. Spence, K. Freeman, A. Morehead, and J.D. Thomas, Real-time Transmission of Full-Motion Echocardiography over a High-Speed Data Network: Impact of Data Rate and Network Quality of Service, Journal of the American Society of Echocardiography, vol. 13, pp. 764–770 (2000). D. Ehler, J.L. Vacek M. Gowda, K.B. Powers, Transition to an All-Digital Echocardiograpgy Laboratory: A Large, Multi-site Private Cardiology Practice Experience, Journal of the American Society of Echocardiography, vol. 13, pp. 1109–1116 (2000). M.J. Garcia, J.D. Thomas, N. Greenberg, J. Sandelski, C. Herrera, C. Mudd, J. Wicks, K. Spencer, A. Neumann, B. Sankpal, and J. Soble, Comparison of mpeg-1 digital videotape with digitized svhs videotape for quantitative echocardiographic measurements, Journal of the American Society of Echocardiography, vol. 14, pp. 114–121 (2001). J.D. Thomas, N.L. Greenberg, and M.J. Garcia, Digital Echocardiography 2002. Now is the time, Journal of the American Society of Echocardiography, vol. 15, pp. 831–838 (2002). B.G. Haskell, A. Puri, and et al., Digital Video: An Introduction to MPEG-2: Chapman&Hall, 1997. Y. Wang, J. Ostermann, and Y.-Q. Zhang, Video Processing and Communications: Prentice Hall, 2002. A. Nicolaides, M. Sabetai, T. Tegos, M. Grifin, G. Geroulakos, S. Danjil, T. Elatrozy, M. Cariolou, E. Bashiardis. Vascular ultrasound, genetic epidemiology and molecular biology: can we bridge the gap?, in R.M. Greenhalgh, J.T. Powell, A.W. Powell (eds), Vascular and Endovascular Opportunities. Saunders, 2000. 93–108. Canny, J.F., A Computational Approach to Edge Detection, IEEE Trans. Pattern Anal. Machine Intell., Vol. 8, pp. 34–43, 1986. Jain, A.K., Fundamentals of Digital Image Processing, Prentice Hall, Englewood Cliffs, 1989. Tabatabai, A.J., Mitchell, O.R., Edge Location to Subpixel Values in Digital Imagery, IEEE Trans. Pattern Anal. Machine Intell., Vol. 2, pp. 188–201, 1984. Lyvers, E.P., Mitchell, O.R., Akey, M.L., and Reeves, A.P., Subpixel Measurements Using a MomentBased Edge Operator, IEEE Trans. Pattern Anal. Machine Intell., Vol. 11, pp. 1293–1008, 1989. Konishi, S., Yuille, A.L., Coughlan, J.M., and Zhu, S., Statistical Edge Detection:Learning and Evaluating Edge Cues, IEEE Trans. Pattern Anal. Machine Intell., pp. 57–74, 2003. Zwet, P., Nettesheim, M., Gerbrands, J., and Reiber, J., Derivation of Optimal Filters for the Detection of Coronary Arteries, IEEE Trans. Med. Imag., Vol. 17, pp. 108–120, 1998.


25

[82] Loupas, T., Mcdicken, W.N., and Allan, P.L., An Adaptive Weighted Median Filter for Speckle Suppression in Medical Ultrasonic Images, IEEE Trans. Circuits Syst., Vol. 36, pp. 129–135, 1989. [83] Karaman, M., Kutay, M., and Bozdagi, G., An Adaptive Speckle Suppression Filter for Medical Ultrasonic Imaging, IEEE Trans. Med. Imag., Vol. 14, pp. 283–292, 1995. [84] Li, C. and Fox, M.D., Analysis and detection of color edges using L2 norm of local covariance matrix of a vector field, submitted to Pattern Recognition, 2003. [85] Edelman, R.R., D. Chien, and D. Kim, Fast selective black blood MR imaging. Radiology, 1991. 181(3): pp. 655–60. [86] Fayad, Z.A., et al., Computed tomography and magnetic resonance imaging for noninvasive coronary angiography and plaque imaging: current and potential future concepts. Circulation, 2002. 106(15): pp. 2026–34. [87] Liu, K. and B.K. Rutt, Sliding interleaved kY (SLINKY) acquisition: a novel 3D MRA technique with suppressed slab boundary artifact. J Magn Reson Imaging, 1998. 8(4): pp. 903–11. [88] Parker, D.L., C. Yuan, and D.D. Blatter, MR angiography by multiple thin slab 3D acquisition. Magn Reson Med, 1991. 17(2): pp. 434–51. [89] Prince M.R., Grist T.M., Debatin J.F., 3D Contrast MR Angiography, Springer Verlag, 3rd Edition, 2003. [90] Wu D.H., Burl M., Reden L., Carlon J.C., MRI RF Coil Systems Having Detachable, Relocatable and/or Interchangeable Sections and MRI Imaging Systems, USPTO.

26


Medical Image Retrieval Based on Plaque Appearance and Image Registration Jaume AMORES and Petia RADEVA Centre de Visió per Computador, Edifici O, Campus UAB, 08193 Bellaterra, Barcelona, Spain Abstract. The increasing amount of medical images produced and stored daily in hospitals needs a datrabase management system that organizes them in a meaningful way, without the necessity of time-consuming textual annotations for each image. One of the basic ways to organize medical images in taxonomies consists of clustering them depending of plaque appearance (for example, intravascular ultrasound images). Although lately, there has been a lot of research in the field of Content-Based Image Retrieval systems, mostly these systems are designed for dealing a wide range of images but not medical images. Medical image retrieval by content is still an emerging field, and few works are presented in spite of the obvious applications and the complexity of the images demanding research studies. In this chapter, we overview the work on medical image retrieval and present a general framework of medical image retrieval based on plaque appearance. We stress on two basic features of medical image retrieval based on plaque appearance: plaque medical images contain complex information requiring not only local and global descriptors but also context determined by image features and their spatial relations. Additionally, given that most objects in medical images usually have high intra- and inter-patient shape variance, retrieval based on plaque should be invariant to a family of transformations predetermined by the application domain. To illustrate the medical image retrieval based on plaque appearance, we consider a specific image modality: intravascular ultrasound images and present extensive results on the retrieval performance. Keywords. Retrieval, contextual information, registration, elastic mathching, medical imaging, IVUS

1. Introduction Images provide a powerful means to represent data, and many applications have as fundamental components the acquisition, processing and storing of huge amount of images. A typical example of such application is medical imaging. In hospitals working with medical images, large amounts of image data are received daily for processing, analysis and archiving. This arises the necessity of constructing database management systems able to organize, analyze and retrieve this type of data. The first image retrieval systems were based on information retrieval methodologies applied to textual annotations. As it has been seen in the last decade, describing images by textual annotations is not satisfactory for three main reasons: i) there is too much information in an image, so that small sets

J. Amores and P. Radeva / Medical Image Retrieval Based on Plaque Appearance

27

of words cannot describe it well; ii) labelling an image by words is a subjective task, as different individuals can apply different labels (words) depending on what they consider more relevant in the current image; iii) a complete manual labelling in a large collection of images is a very tedious task. These reasons have led to the use of image content for processing and organizing the database, constituting the so-called content-based image retrieval systems. Such content-based image retrieval (CBIR) systems are necessary for a wide range of applications dealing with vast volumes of images. In the medical image field, retrieval by content is necessary for performing guided diagnosis and therapy. Having medical images archived along with their descriptions, retrieval by content allows the physician to present a query image representing the current clinical case, and obtain the most similar stored images and their associated descriptions, so that the diagnosis of the image at hand is more easily taken by comparison. Indeed, whenever the type of the medical image is of great complexity it is common for the physician to base its diagnostics on manuals containing for each pathology its representative images. Other applications are the construction of electronic atlases with a high number of examples for each case, and educational uses. In this chapter, we overview the work on medical image retrieval, underline the peculiarities of medical image retrieval as opposed to image retrieval of general domains, and present a general framework of medical image retrieval based on plaque appearance. A fundamental aspect of any CBIR system is the use of an appropriate feature space. This feature space should be able to represent all aspects of the image relevant to its description. For doing so, not only global information, but also local and contextual information is necessary. Global information is necessary for describing the statistics of relevant features inside the image. Local information is fundamental for dealing with images in which important parts are localized in small regions such as the pathology bearing regions of medical images. In addition to local and global information, the feature space should also be able to describe the spatial relations of the different structures conforming the overall image. In retrieval of general scope, this contextual information is good for describing semantically complex scenes as the relation between the objects conforming them. In the medical image retrieval field, the relative disposition of pathological structures respect to other structures play key roles in diagnosis. Finally, in complex domains such as retrieval of medical images the design of the feature space must be flexible enough to incorporate descriptors specific to the concrete problem, as general descriptors perform poorly in describing them. 1.1. State of the Art in General Content-Based Image Retrieval and Medical Image Retrieval Most of the general CBIR systems up to day try to characterize the whole image by using only global information, i.e. a set of global signatures such as histograms of color, texture or shape [30,13]. Pentland et al. ([24]) use different global descriptors depending on the type of the image. Their global description is based on an orthogonal feature space representing the content, using eigenimages for objects such as faces, eigenmodes for retrieving shapes and the world-based features for textured images. In retrieval of general scope, some authors [28,7,31,9] have included recently local information able to differentiate between images in which an important part of the discriminant information

28


is localized in small regions. Schmid et al. ([28]) extract characteristic points localized in discriminant parts, and compute a set of invariant local features around each characteristic point. Carson et al. [7] segment the images using expectation–maximization based on texture and color. They achieve weak segmentation: the image is segmented into blobs which may represent objects as a whole, parts of objects, or unions of different objects. This weak segmentation, although does not represent isolated objects, approaches the semantics by which users organize images, and permits better retrieval results when the user queries for discriminant objects in the image. Wang et al. [31] computes a weak segmentation based on k-means clustering of color and texture features. Their algorithm is faster than the used by Carson in [7], but achieves more inaccurate segmentations. They propose a similarity measure among images that compensates inaccurate segmentations by allowing multiple matchings from regions of the query to regions of the target image. Chen et al. [9] following the same idea compensates inaccurate segmentations by treating the segmented regions as fuzzy sets and using a unified fuzzy matching between regions. Contextual information or (relative) spatial layout have also been explored by some authors. A typical way to deal with spatial organization is generalizing the concept of string to a plane, conforming the so-called 2D string descriptor [8] and 2D C string [17]. In this descriptor, objects are represented by nodes and their spatial relationship by strings. 2D strings requiere very accurate segmentations of the objects in the image, what leads in practice to manual segmentations. Another descriptor for describing context is the correlogram. A correlogram is an histogram measuring the distribution of features such as color in the image as well as their spatial relationship (or their co-ocurrence). Huang et al. use these correlograms in retrieval of images by color, and Belongie et al. use another type of correlogram, called shape context, for taking into account the statistics of the distributions of points in shape matching and retrieval. These descriptors are good for taking into account the spatial distribution of the colors [15] or the spatial distribution of binary shapes [4], but do not permit to take into account context of different structures. We will show how generalizing them we can have this contextual information without needing manual segmentations. The extension of image retrieval to medical image applications is a challenging and emerging field where few works have been published. Y. Liu et al. [19] characterizes retrieval as a classification problem. This is correct when the images can be categorized as belonging to known classes (e.g. representing different diseases in medical diagnosis) and the user wants to retrieve images that belong to the same class as the example presented. Following this idea the authors begin with a big set of descriptors and find the set of weights that achieves the lowest classification error. They basically use global descriptors specifically chosen for CT brain images, exploding the symmetry in the brain as a basic property of normal (not diseased) brains. Local information is also extracted for asymmetrical regions. Other authors such as Korn et al. [16] explore the use of fast spatial access methods such as R-trees and fast nearest neighbor search. Their work is based on artificial sets of data that simulate tumor-like shapes. Focusing on the design of appropriate feature spaces, Kak and Brodley [29,11] take local and specific information for each of the pathology bearing regions (PBR) previously delineated by the user (the physician). Their work is specifically applied to highresolution computer tomographies (HRCT) of the lungs. For each image, low-level features are extracted locally at every region and globally for the whole image. They use an


29

exhaustive set of descriptors that includes the usual features based on texture, gray-level and shape, but also includes very specific descriptors designed to describe the different lung diseases and their appearance in high-resolution computed tomographies (HRCT). Finally, they take only the most relevant information by sequential forward selection search, which reduces the dimension of the feature set. Very specific contextual information is also extracted by recording the relative position of the manually segmented region respect to manually extracted fissures of the lung ([29]). The high cost of the manual delineation makes it impractical for regularly actualized large image collections. Liu and Sclaroff [18] perform segmentation of blood cell micrographs based on deformable shape models. The image is over-segmented and candidate regions are then merged into whole objects by using a deformable shape model and the regions color compatibility. Finally, they perform population-based retrieval of cells. For every image, an histogram of the shapes of the cells in the image is computed, and retrieval is achieved by shape histogram similarity. Paredes, Lehmann et al. [23] represent each image by several square windows which may overlap, and use them as a set of local appearances for the image. A condition for robust retrieval using local appearance is the preservation of shape. Regarding contextual information P. Liu et al. [14] use the centers of mass of manually segmented objects and their geometric relations to deal with the contextual information in retrieval of magnetic resonance images (MRI) of the chest. Petrakis et al. [25] use graphs, for retrieval of MRI images of the brain. Both 2D strings used by Chang et al. [8] and graphs used by Petrakis [25] permit a flexible form of describing any type of image in terms of its structures (including their local attributes) and their spatial relations. The authors achieve in these works highly discriminative descriptors that can be indexed efficiently when the images have structures segmented manually, but not robust for automatic (non exact) segmentations. In addition to having an appropriate feature space, it is highly desirable to have invariance to a particular family of spatial transformations determined by the application domain. For achieving this we can employ either invariant feature spaces or similarity measures that are not affected by these spatial transformations. This invariance is very important in medical images, as there is a high degree of shape and appearance variability inter and intra subject. Despite this fact, there are few works in retrieval that deal with this type of invariance. In medical image retrieval, Lehmann et al. [10] use a distance measure invariant to small global transformations, and a distortion model that compensates more local deformations. This distortion model allows each pixel to be matched to any pixel of the destination image around a local neighborhood, without guaranteeing any regularity or topology preservation in the mapped object. Tagare et al. [27] compute the similarity between pairs of shapes by first registering them. They forbid changes in topology by setting a set of constraints in the matchings, and solve the combinatorial problem by dynamic programming. Their method, however, is only valid for matchings between contours, and does not attempt to achieve any smoothness in the mapping. Y. Liu et al. [20] make retrieval of 3-D CT brain volumes by first registering them, using a specific method based on the symmetric properties of the brain and an affine transformation, non invariant to local elastic deformations. In this chapter we present a general framework of medical image retrieval based on plaque appearance, and consider a specific image modality: intravascular ultrasound images.

30


Figure 1. IVUS image with different structures.

1.2. Medical Image Retrieval Based on Plaque Appearance, Application to IVUS An IntraVascular UltraSound (IVUS) image is obtained by inserting a catheter into the artery with a transducer on its tip. The transducer emits ultrasonic waves as it rotates, and these waves are propagated according to the physical properties of the media the catheter is in. Depending on the echogenic impedance of some structure, and its proximity to the tip, the wave will be reflected with some amplitude and some delay, which allows to form a cross-sectional image of the artery: the delay accounts for the distance of the structure from the tip, and the amplitude accounts for the intensity of the gray level used for representing the structure in the image. Figure 1 shows an IVUS image with two calcium plaque structures (regions of high level with a shadow behind), the catheter, and the usual adventitia tissue. The shadow behind the calcium plaques is due to the high impedance of the calcium, which does not allow the ultrasonic wave to pass through. We refer to [12] for an introduction on the topic. IVUS images of the coronaries are a novel and, at the same time, key tool in the correct diagnosis of coronary diseases such as atherosclerosis, which can provoke heart attacks. The great difficulty of these images makes it particularly interesting for the physician to perform analysis based on a history of similar cases, which motivates the construction of CBIR systems of IVUS. This CBIR should take into account the spatial relative distribution of structures such as atherosclerotic plaques, as studies have shown ([21]) that spatial characteristics such as the size of these atherosclerotic plaques, their eccentricity, and degree of embracement around the vessels; play important roles in the study of heart diseases. Furthermore, the intrinsic elastic properties of the arteries make these objects have an extraordinary high variation intra-subject and inter-subject, which demands invariance to elastic transformations in retrieval similarity between IVUS images. 1.3. Context Based Plaque Retrieval In contrast to the discussed approaches, the Context Based Plaque Retrieval method uses a feature space that integrates all types of information important in describing the content of the image: local, contextual and global information. This makes it appropriate to deal with medical images that hold complex information. For doing so we make a general-


31

ization of the correlograms to deal with images presenting different types of structures. Correlograms are histograms that take into account not only the statistics of the features in the image, but also the relative spatial distribution of these features. Using this generalization of the correlograms to conform our feature space has various advantages: First, it incorporates all the types of information mentioned above. Second, it does not need accurate segmentation/classification of the structures inside the image. This fact permits to make an automatic segmentation of the structures instead of making manual segmentation of the image, as opposite to the majority of descriptors used for dealing with this contextual information. Finally, the generalized correlograms permit easily to incorporate local information specific to the image application domain, which is mandatory in complex domains such as medical images [8,25,14]. For dealing with invariance to elastic transformations, an efficient elastic matching method is proposed, aligning each pair of images before their comparison. This is achieved by using a sparse set of landmarks placed around salient regions of the image, and computing a fast transformation such as the thin-plate spline based on matching between these landmarks. The thin-plate spline produces elastic alignments modelling both global and local deformations, and restrains the object from undergoing unnatural deformations (e.g. changes in topology) by performing regularization. In the feature space component of the registration, we also make use of the proposed generalized correlograms. This allows us to match structures attending not only to their local properties (the type of the structure), but also to the global attributes (such as their size) and their context. Furthermore, these correlograms introduce spatial coherence and remove ambiguities in the computation of the correspondences, which makes the algorithm achieve good solutions with few iterations.

2. Feature Space As mentioned before, it is important that the feature space takes into account all types of information relevant to retrieve the image. Thus, we include local, global and contextual information, and we do so by using generalized correlograms. 2.1. Local Information By using local information we aim at describing the different types of structures inside the image. In the IVUS case, the discriminating structures are placed around the wall of the vessel [12]. A snake is placed at the center of the image after applying an anisotropic diffusion, and it is attracted to the wall. The set of landmarks is then obtained by sampling the snake (see Fig. 2). Associated to each landmark, a local feature vector is computed that describes the type of structure where the landmark lies. Here is where we include specific information about our domain, as these local feature vectors must be specifically chosen for characterizing the structures of our particular medical domain [11]. In our IVUS case, we take the gray level profile along the normal to the wall at the landmark, in the direction from the landmark to the outward part of the vessel (Fig. 2). This descriptor is specially appropriate for characterizing IVUS structures, see [1] for further details. Finally, these local feature vectors are classified and labels are assigned to each landmark, giving more compact local information. In our case non-parametric discriminant analysis [5] is performed followed by K-NN classification [1].

32


Figure 2. Landmark and local feature vector extraction.

2.2. Global and Contextual Information We incorporate this local information into a generalization of correlograms that allows to provide this information along with contextual and global information about the image. Correlograms are histograms which not only measure statistics about the features of the image, but also take into account the spatial distribution of these features. For doing so a spatial quantization of the points inside the image must be done. Regarding this spatial quantization we provide two different definitions of the correlogram: a bidimensional definition and a periodic unidimensional correlogram. The bidimensional definition takes the same spatial quantization of the shape context descriptor of Belongie et al. [4]. The proposed generalized correlogram is then defined by adding a dimension to the correlogram. This dimension takes into account the types of structures of the landmarks around the one being described. Let C be a set of n landmarks, and pi ∈ C the current one being described. Let lj be the label of the type of structure where the landmark pj lies. Let nc be the maximum number of classes. For pi the correlogram hi is defined as: hi (r, θ, c) = #{pj ∈ C, pj = pi : (pj − pi ) ∈ Dr , (p j − pi ) ∈ Aθ , lj = c},

(1)

where Dr is the r-th interval of radius: r = 1 . . . nr , Aθ is the θ -th interval of angles: r θ = 1 . . . nθ , and c represents the class: c = 1, . . . , nc . The sets of intervals {Dr }nr=1 and nθ {Aθ }θ=1 constitute a spatial quantization of the possible angles and possible distances of the relative positions around the current landmark. Thus, the first two dimensions of the generalized correlogram take into account the relative spatial position of the points (or landmarks), and the third dimension takes into account their type of structure. The correlogram can be interpreted then as an histogram measuring the density over relative positions of the different types of structures. The landmarks whose distance and angle relative to pi lie at Dr and Aθ constitute a cell in the plane (see Fig. 3(a)). The size of these cells is incremented exponentially as the positions move away from pi , so that more importance is given to local context. Figure 3(a) shows a bidimensional correlogram applied to a landmark of an IVUS image. The landmarks in this image have been classified into two classes: those belonging to calcium plaque (red points), and those belonging to adventitia (blue points). This correlogram has 12 intervals of angles (nθ = 12)


(a)

33

(b) Figure 3. Bidimensional correlogram.

and 5 intervals of radius (nr = 5). Figure 3(b) shows a log-polar representation of the correlogram for each of the values of the third dimension: type of structure c = 1 and c = 2. In this plot, cells with a high density of points from a particular type of structure are represented by a high gray level. This correlogram has scale and orientation invariance by normalizing the distances pi − pj by the size of our object and orientating the correlogram along the tangent of the shape (see [4]). As said above, this bidimensional definition of the correlogram is a generalization of the one used by Belongie, by extending to take into account landmarks from different types of structures. At the same time, this spatial quantization defined by Belongie can be regarded as a generalization of the one used by Huang et al. [15], as the latter takes only into account distances in the spatial organization. The main disadvantage of this bidimensional quantization is that the resulting correlograms are not robust against large shape changes of the object. This low robustness is accused before registering the images, however it can be avoided by using an appropriate feedback scheme (see below). Still, we make another definition of a correlogram by taking another type of spatial organization, robust to shape changes. This second definition represents a change of manifold from all the plane to the closed curve where the landmarks are placed, as in our case all the landmarks lie along the wall of the vessel. Now the landmarks pi are represented as the arc-length position inside the curve. Let : [0, 1) → R2 be the arc-length parameterized curve. The representation of the i-th characteristic point pi is now taken as its arc-length parameter ui inside this curve: (ui ) = pi , ui ∈ [0, 1). Let C = {ui }ni=1 be the set of arc-length parameters corresponding to the set of landmarks C. The correlogram for the i-th characteristic point is defined as: hi (s, c) = #{uj ∈ C , uj = ui : (uj − ui ) ∈ Is , lj = c},

(2)

where Is is the interval of arc-length positions of the s-th cell, s = 1 . . . ns . By this definition, we are quantizing the possible values of arc-length differences between points of the curve. This difference must be computed with arithmetic modulus 1, in order to take into account the closed nature of the curve. Figure 4 shows a diagram of the spatial quantization of the 1D correlogram. As can be seen the cells form a partition of the closed curve where the landmarks lie. The picture shows cells of equal size for clarity

34


Figure 4. Unidimensional correlogram.

purposes, although the size of the cells is incremented exponentially from pi to outwards, as occurred with the bidimensional correlogram. The advantage of this new spatial definition is that it is robust against changes of shape. The contextual information is now taken along the curve, without being affected by its changes in curvature. The main advantage of using correlograms is that they are robust against missclassifications, so that if some landmarks are missclassified the spatial distribution of the structures around pi is still represented. They also allow to include specific descriptors representing the types of structures in an easy way. We only have to choose a good descriptor for the structures of a particular medical domain, and classify the landmarks according to this descriptor. Then the labels are included in the correlograms as explained before. We also use generalized correlograms as a descriptor of the landmarks in the registration step. These contextual descriptors produce matching between structures attending not only to their type (local attributes) but also to their global attributes (such as their size) and their context. This solves ambiguities in the possible correspondences for each landmark. It also provides spatial information leading to regularity and coherence in the set of correspondences, accelerating the registration process. In this work we provide results when using bidimensional and unidimensional correlograms. Both of them are valid for registering the images, although as we will see unidimensional correlograms are more efficient than bidimensional correlograms. The disadvantage of the former is that it is only valid when the landmarks form a closed curve. All along we have defined the correlograms to measure the density over relative positions of the different types of structures around pi . We can also define an autocorrelogram, as called by Huang et al. [15] which takes into account only how the points from the same type of structure are organized around pi : hi (r, θ ) = #{pj ∈ C, pj = pi : (pj − pi ) ∈ Dr , (p j − pi ) ∈ Aθ , lj = li )} An analogous definition is done using an unidimensional definition as Eq. (2). Autocorrelograms are used just for registration purposes. Given two sets of landmarks from two structures of the same type, using auto-correlograms we produce matchings between points in the same relative position inside the structure, e.g. matching extremum points together, middle points together, and so on. This auto-correlogram is used in a refinement step, once we have two structures of the same type and with the same context coarsely aligned. The aim is to make correspondences between homologous structures more exact and regular, discarding the information of the rest of structures.


35

Figure 5. Scheme of the overall feature space design.

Finally, we can use fuzzy definitions of the correlograms defined before. Until now each point pj lying in some cell adds 1 to the count of this cell, in the dimension lj = c. We can add to the count in the dimension c the probability P r[class(pj ) = c]. In this work, however, we do not use fuzzy definitions. In Fig. 5 we can see a scheme illustrating the whole feature space construction. Finally, regarding the similarity measure between correlograms, we use the χ 2 statistics to compare them: 1 (hi (k) − hj (k))2 2 hi (k) + hj (k) d

χ 2 (hi , hj ) =

k=1

where d is the dimension of the correlogram, if we treat it as a unidimensional vector (e.g. in the case of the definition of Eq. (1), we arrange the matrix of dimension nr × nθ × nc into a vector of one dimension d × 1, d = nr nθ nc ). This distance acts like the Euclidean metric normalized by the number of points falling in the k bin of both histograms hi and hj . It has been used among others by Huang et al. [15], and Belongie et al. [4].

3. Registration: Obtaining Invariance Against Elastic Transformations We obtain invariance against elastic deformations through registering the images before their comparison. The scheme followed in the registration is the so-called point-mapping (Fig. 6). First a set of landmarks is extracted from each image. The landmarks are described in some feature space, and a set of correspondences is computed which globally minimize the distance between landmarks in this feature space. Finally, a transformation is obtained based on these correspondences. The proposed registration method also includes a search strategy of the final transformation. The feature space is the same as the explained in the previous section. 3.1. Computing Correspondences Once described the characteristic points in the feature space we compute the distance between any characteristic point in the image I1 and any characteristic point of the image I2 and based on this distance obtain our set of correspondences. We let for Section 3.3 the

36


Figure 6. Scheme of the point-mapping paradigm.

expression of the exact distance. Let d(i, j ) be the distance between landmarks pi ∈ I1 and qj ∈ I2 . We want to obtain a correspondence function φ : {1, . . . , n} → {1, . . . , n} that minimizes ni=1 d(i, φ(i)). Such a function can be obtained by an assignment optimization algorithm such as the Hungarian’s method [22]. 3.2. Finding a Transformation Once we have our set of correspondences we want to find a transformation function T : R2 → R2 , mapping the coordinates of I1 onto I2 . This function must map the characteristic points of I1 close to their correspondent positions of I2 , and produce a smooth mapping for the rest of points of I1 . Any registration algorithm is characterized by the family of transformations to which it provides invariance. Biological bodies have a great deal of variability, making necessary the use of elastic transformations. Elastic matching methods allow to model global changes or transformations and local (elastic) deformations. Furthermore, they lead to smooth (regular) transformations avoiding changes in topology of the object. For doing so, the transformation must incorporate a regularization component. In the present work we study the use of Thin-Plate Splines (TPS) as elastic matching method. TPS is an efficient method when based on a small set of landmarks, as the computation of the transformation is done by a closed-solution formulae which involves inverting one matrix of n × n elements, n being the number of landmarks. A Thin-Plate Spline based transformation for an image can be expressed as Tλ (x, y) = (f x (x, y), f y (x, y)), where f x (x, y) and f y (x, y) are two independent surfaces obtained from the set of correspondences, and λ is the regularization parameter of the transformation. Let {xi , yi }ni=1 be the n landmarks of the origin, and {xi , yi }ni=1 the corresponding landmarks at the destination: f x (xi , yi ) = xi , i = 1, . . . , n and f y (xi , yi ) = yi , i = 1, . . . , n. We explain how to obtain one of the surfaces, let f (x, y) be such a surface. We call {ti }ni=1 the known data for the function we want to interpolate, so that we restrain f (xi , yi ) = ti , i = 1, . . . , n. Hence, if the surface f to estimate is f x , then {ti }ni=1 = {xi }ni=1 . If the surface f to estimate is f y , then {ti }ni=1 = {yi }ni=1 . Then f (x, y) can be expressed as:


f (x, y) = d1 + d2 x + d3 y +

n

ci k (x, y) − (xi , yi )

37

(3)

i=1

where k is the basis function or kernel: k(r) = r 2 log(r), k(0) = 0. The coefficients d1 , d2 , d3 , c1 , . . . , cn are obtained solving a system of linear equations derived from applying the following restrictions: i) interpolation conditions (n equations): f (xi , yi ) = ti , i = 1, . . . , n; ii) in order for f (x, y) to have square integrable second derivatives we require: n

ci = 0,

i=1

n i=1

ci xi = 0,

n

ci yi = 0

i=1

which produces a total of n + 3 equations expressed by the following matrix notation: K P c t = (4) d 03×1 P T 03×3 where Kij = k((xi , yi ) − (xj , yj )), the ith row of P is (1, xi , yi ), c = (c1 , . . . , cn )T , d = (d1 , d2 , d3 )T , t = (t1 , . . . , tn )T and 03×1 denotes a 3 × 1 matrix of zero elements. Let L be the (n + 3) × (n + 3) matrix of the last equation, a the vector of coefficients a = (c, d)T and v = (t, 03×1 )T . We need only to invert the matrix L for obtaining the coefficients of Eq. (3), a = L−1 v, and as discussed in [26] this matrix is non singular. The steps described above obtain an interpolating surface f , for λ = 0. If we want to obtain a regularized (i.e. approximating) surface, we just have to substitute in Eq. (4) the matrix K by K + λIn×n , where In×n is the n × n identity matrix [2]. After applying TPS we can also obtain a measure of the bending energy this transformation has. Let ax be the vector of coefficients a, defined above, for component x of the transformation. Let K be the kernel matrix defined also above, then the bending energy due to the component x of the transformation is: Ex = axT Kax

(5)

Let Ey be the energy employed in the y mapping f y , obtained in an analogous manner. We have as energy E of the whole transformation the sum of both components: E = Ex + Ey . Finally, the regularity of the TPS for different values of λ depends on the square of the scale of our objects. So if we have s as the size of our objects, we should compute λ = λnorm s 2 , and experimentally set values for λnorm on normalized objects. The λ parameter represents the tradeoff between approximation to the data (the computed correspondences) and smoothness. A high value of λ smoothes out irregularities in the correspondences, necessary in the first steps of the algorithm. At the same time, a high λ leads to a more global (coarse) transformations, whereas a low λ makes the transformation more elastic. The transformation can then accommodate more local deformations, refining the alignment. The regularity of the set of the correspondences depends on the stage of the algorithm. Initially they are quite irregular due to the complexity of the problem, and they become more and more regular through the feedback scheme we will explain below.

38


Finally, the computed transformation also permits to produce a warping of image I1 , making it more similar to image I2 . This is necessary for taking into account the appearance similarity in the final comparison (see Section 4). 3.3. Search Strategy Many registration algorithms [6] include some search strategy of the final transformation. We include this component through a simple yet effective feedback scheme performed in an iterative manner. The complexity of the images leads to irregular nonreliable initial sets of correspondences. However, the correspondences do obtain the correct global information to align coarsely the images. Thus, we introduce feedback of coarse transformations to obtain finer ones. The feedback is made by biasing the next set of correspondences towards the positions obtained by the last transformation. This is done by recomputing the distances between landmarks pi ∈ I1 and qj ∈ I2 as: dij = dijF + αTλ (pi ) − qj , where dijF is the distance between landmarks in the feature space, and α is the degree of influence of the last transformation in the next set of correspondences. This feedback achieves a simultaneous maximization of the regularity of correspondences and the similarity of corresponding points throughout the iterations. The regularity is enforced by the term αTλ (pi ) − qj , that represents the difference of the vector formed by the potential correspondence u = (qj − pj ) from a regular correspondence v = (Tλ (pi ) − pi ): v − u = Tλ (pi ) − qj . The similarity is enforced by the term dijF . It can also be seen as a form of cooperation, as defined by Brown in [6]. In a cooperative scheme the correspondences computed for neighbor landmarks give information about the correspondence computed for the current landmark. Here the TPS mapping for the current point, Tλ (pi ), depends on the obtained correspondences in its neighborhood, and we restrict a potential matching point qi not to lie far away from this mapping. The α and λ parameters representing the influence of the last transformation and the degree of regularization are updated throughout the iterations following an annealing scheme. The α parameter represents the confidence in the last transformation, whereas a low λ parameter represents confidence on the last set of correspondences (i.e. correspondences without many irregularities). As both the correspondences and the transformation are improved throughout the iterations, we must increment α and decrement λ. Both are changed with an exponential ratio, see [1] for further details. Finally, we can consider part of the search strategy the hierarchical scheme followed in the computation of global to local transformations. Furthermore, a hierarchical scheme is followed in the use of global to local information, as we use bigger correlograms for computing the first transformations (global alignments require more global information) and smaller correlograms for making more fine alignments.

4. Similarity Measure in the Final Comparison Between Images The registration produces as output a function T which is regular and maps the characteristic points pi from I1 close to their corresponding ones in I2 . However, the mapped points are not exactly the characteristic points qi of I2 . For obtaining a regular final set of correspondences φ from {pi }ni=1 to {qi }ni=1 , we simply take the Euclidean distances of


39

mapped points and destination points: dij = T (pi ) − qj and compute the correspondences using the hungarian’s algorithm over this matrix of distances. Our similarity measure is based on the sum of three factors: the distance in the feature space described earlier, the amount of deformation necessary to align both objects, and a local appearance difference between the aligned image and the destination image. The amount of deformation is measured through the energy E of the Thin-Plate Spline computed over the regular correspondences obtained before. A high energy means that the amount of deformation is too high for considering natural this transformation. The distance in the feature space is in our case the distance of the correlograms. We can use either the 2-D correlograms or the 1-D correlograms. We have seen small difference between them in the results because the 2-D correlograms perform as well as the 1-D correlograms when the shapes are already aligned. As 2-D correlograms are more general, we used it. We recompute these correlograms orienting them now along the x axis of the image. This is done to reduce the little robustness that these correlograms have on the computation of the tangents, now that the alignment permits us to avoid the invariance to orientation. The size of the correlograms is the maximum size in the hierarchical approach, so that all the image is included in every correlogram. Let dijF be the χ 2 distance between correlograms of pi ∈ I1 and qj ∈ I2 . We compute the global distance between both images, d F (I1 , I2 ) as the symmetric sum over distances of best matching correlograms: d F (I1 , I2 ) =

1 1 arg min dijF + arg min dijF j i n m n

m

i=1

j =1

For computing the local appearance difference between both images, let IW be the image I1 warped according to the obtained transformation from the registration. We take local windows around the mapped points T (pi ) in IW and matching points qφ(i) in I2 . The local appearance difference is expressed as: d A (I1 , I2 ) =

w n w

2 G (x, y) IW (T (pi ) + (x, y)) − I2 (qφ(i) + (x, y))

i=1 x=−w y=−w

where G(r) is a gaussian-like function of the radius r, more sensitive to close positions. The warped image IW does not respect the original pattern of the textures, so it is better to remove them in the comparison. Therefore, we take as images I1 and I2 the anisotropic diffusion of the original images. Finally, the total distance between both images is computed as a combination of the distance components defined above: d(I1 , I2 ) = α F d F (I1 , I2 ) + α A d A (I1 , I2 ) + α E E This distance is used by Belongie et al. in [4], for binary objects. We apply it generalized for objects with several structures. The weights α F , α A , α E are computed as the ones minimizing the classification error on the IVUS database, following a leave-one-out procedure.

40


Figure 7. Histogram of the efficiency of classification.

5. Results In this section we evaluate the system in three aspects: i) the rate of success of the local classification of tissues plaque/versus non-plaque, ii) the results of applying each of the components of the registration algorithm: the proposed feature space, the feedback scheme, and the overall registration results; and finally, iii) the overall retrieval results. All the experiments have been conducted on a database of 100 IVUS images, all of them presenting plaque structures. Studying the registration of images presenting plaques is of high interest for the following reasons: first, there is a great difficulty in the differentiation between plaques and adventitia tissue, second, there is a high variability in the shapes of both the entire vessel and the plaque structures, and third it has been clinically seen that the relative spatial distribution of the plaques is important in diagnosis of heart diseases. This makes it interesting to study the use of the proposed contextual descriptors on these images in order to retrieve them. As explained in the introduction, retrieval of medical images is very interesting to perform computer-aided diagnosis. In particular, retrieval of IVUS images is useful to assist coronary diagnosis and intervention. 5.1. Local Classification Results As explained in Section 2, when building the feature space, we first obtain a set of local descriptors based on gray level profiles along the normal to the contour at this point, and then perform classification of these feature vectors, assigning labels to each characteristic point. We have designed a classification algorithm based on reducing the dimensionality by Non-Parametric Discriminant Analysis [5], and nearest neighbor classification. The scope is to achieve moderately good classification results. As the classification efficiency is different depending on the particularities of each image, we evaluate the efficiency of classification for all the landmarks in each image, and compute the mean efficiency and standard deviation of classifications across all the images. In Fig. 7 we show the histogram of the efficiency of classification, the horizontal axis represents the efficiency of classification and the vertical axis the ratio of images that whose landmarks are classified with this efficiency of classification. The red line indicates the mean efficiency, which is 90.2%, and the green lines indicates the mean minus and plus the standard deviation. The standard deviation is of 7.8%. We must note that the scope of this work is neither to design a very accurate local specific feature, nor to employ a sophisticated classification scheme over this local feature space. The results achieved in classification are good enough to work with


(a)

(b)

(c)

(d)

(e)

(f)

41

Figure 8. Some classification examples. Red points indicate classification as calcium plaque, and blue points as adventitia (the rest of the tissue). Correct calcium structures are indicated by green boxes surrounding the structure. (a–b) good results, (c–d) regular results, (e–f) bad results.

the proposed correlograms for performing registration and retrieval. In Fig. 8 we show some classification examples. Red points represent landmarks automatically classified as belonging to plaque structure, whereas blue points represent landmarks belonging to non-plaque tissue. With cyan boxes we have indicated the correct regions where landmarks belong to plaque structure, in order to show the goodness of the automatical classification. The majority of images has good classification results, as in the case of Figs 8(a)–(b). The worst examples are shown in Figs 8(e)–(f). The main problem is that

42


in some images compact regions are wrongly classified as belonging to plaque structure, and therefore, these regions are interpreted by the correlogram as structures (false structures). This is difficult to avoid because there are some regions of adventitia resembling much to plaque. These are regions of high gray level and with a black region behind it, which is what characterizes mostly calcium plaque. Figures 8(e)–(f) show such regions. 5.2. Performance of the Correlograms In the registration algorithm, the use of correlograms has several advantages: first, it allows to obtain more regular sets of correspondences, as it introduces spatial coherence information in the feature space. Second, it incorporates not only contextual information, but also global characteristics of the different types of structures. This produces matching of homologous structures having the same global characteristics such as size. We first show how correlograms obtain regular correspondences. Figure 9 shows a couple to be registered. Both images have only one plaque structure, the homogeneous white structure. Figure 10 shows correspondences obtained by different descriptors for the couple of Fig. 9. The original landmarks are placed in a small region of adventitia in the left image of each pair. Figure 10(a) shows the correspondences when using local feature vectors. The set of correspondences does not hold spatial regularity as spatial information is not included in the descriptor. Figure 10(b) shows the correspondences when using local gray level windows of size 19 × 19. Taking local windows is not robust against shape changes, and gray-level windows are poor in IVUS. As can be seen, the set of correspondences is not regular. Figure 10(c) shows correspondences using 2D correlograms. The regularity is not complete, but is much higher than with the other descriptors. Finally, correspondences using 1D correlograms are shown in Fig. 10(d). The set has very few irregularities, and it is more regular than using 2D correlograms. Now we see the effect of using correlograms for taking into account global characteristics of the structures. Figures 11(a)–(b) show a new couple of IVUS images to be registered, (a) displays the image I1 to be aligned and (b) the destination image I2 . Figures 11(c)–(d) show the anisotropic diffusion of (a)–(b) respectively. In red we superpose the contour of the vessel from which the characteristic points are extracted in each image. The image I1 has two calcium plaques on both sides (indicated in the Fig. 11(a)), and the image I2 has three calcium plaques: two on both sides and one at the bottom (indicated in the Fig. 11(b)). Taking into account global characteristics the plaques on both sides should be matched in both images, leaving alone the small plaque at the bottom of the image I2 . We show how the global description is included with correlograms by comparing the result of a first coarse transformation using contextual information (2D correlograms) and then using only local information (the local feature vectors). We show transformation results on the anisotropic diffusion of the images because it is visually more clear. Figure 12(a) shows the warped I1 when using correlograms. Figure 12(b) shows the image I2 with the edges of the warped I1 superposed in red. We can see how both calcium plaques of I1 are mapped close to the big plaques of I2 , as well as the adventitia tissue. Figures 12(c)–(d) show the warping result when using only local feature vectors. One of the calcium plaques (indicated in Fig. 12(e)) has not been mapped close to any of the plaques of I2 , as the mapped plaque (indicated by a green arrow in Fig. 12(d)) lies at an intermediate position between a big plaque and a small one (blue arrows). Using correlograms this matching is avoided as the size characteristic is included.


(a)

(b)

(c)

(d)

(e)

(f)

43

Figure 9. First couple example to be registered. (a) I1 image to be aligned, (b) I2 destination image; (c)–(d) anisotropic diffusions of (a)–(b) respectively; (e)–(f) Classification results for (a), (b): in red the landmarks classified as belonging to calcium plaque, in blue landmarks classified as belonging to adventitia tissue.

5.3. Evaluation of the Feedback Scheme The result of applying the feedback scheme is that the transformation becomes more and more approximate and at the same time the set of correspondences becomes more and more regular. This is illustrated first qualitatively for the pair of images shown in Fig. 11,

44


(a)

(b)

(c)

(d)

Figure 10. Correspondences using different descriptors (see text).

(a)

(b)

(c)

(d)

Figure 11. Couple of images to register (a)–(b). Their anisotropic diffusions (c)–(d).

and then quantitatively by showing the evolution of the error through the iterations of the algorithm applied to 100 random registrations. Figure 13(a) shows the initial set of correspondences in the registration of the pair of images of Fig. 11, and Fig. 13(b) shows the final set of correspondences. The initial set is very irregular, but holds information about the correct global matching. Figure 13(c) shows the transformation based on this set, where the plaques are placed with the correct orientation and position (see green arrows in right image of (c)). Figures 13(d) and 13(f) show respectively an intermediate and final stage in the computed transformation. The intermediate transformation is coarse (rigid) but more approximated than the initial transformation, whereas the final transformation is accurate and elastic. Now we see a quantitative evolution of the approximation error and degree of irregularity of computed correspondences throughout the iterations of the algorithm. We have made 100 random registrations, and computed the median of the approximation error and irregularity at the successive iterations of the algorithm. We measure the approximation error for an obtained transformation T as:


(a)

(b)

(c)

45

(d)

Figure 12. Coarse alignment (first step of the algorithm) using first contextual information (a)–(b), and then only local information (c)–(d).

(a)

(b)

(c)

(d)

(f) Figure 13. Evolution through the different stages of the algorithm.

Ea (I1 , I2 ) =

Eai (p, C)

nc 1 1 max 1 nc n i i=1

=

min

q∈C,class(q)=i

p∈C1 ,class(p)=i

T (p) − q ,

Eai (p, C2 ),

1 n2i

Eai (q, C1 )

q∈C2 ,class(q)=i

(6)

where I1 , I2 are the registered images, nc is the number of classes (types of structures) we deal with, n1i is the number of landmarks from I1 belonging to class i, and n2i is the same for image I2 , T is the obtained mapping, and Eai (p, C) measures the distance of the mapped point p to points of C from the same class.

46


(a)

(b)

Figure 14. Evolution through the iterations of the approximation error (a) and degree of irregularity (b) using 1D and 2D correlograms. The solid line with squares display the 1D correlogram case, and the dashed line the 2D correlogram case.

We measure the irregularity of each set of correspondences performing a TPS over this set and measuring its bending energy (see the Section 3.2). This measure is high whenever there are irregularities. Figure 14 shows a graphic of the median evolution of the approximation error (Fig. 14(a)) and irregularity (Fig. 14(b)) of the set of correspondences throughout the iterations of the algorithm. We compute this median over 100 randomly chosen pairs of images to be registered. Each pair consists of two images from the same category. In blue dashed line we have this evolution when using bidimensional correlograms, and in red solid line with squares when using unidimensional correlograms. Both using bidimensional and unidimensional correlograms, the approximation error and the irregularity are decreased throughout the iterations. This is due to the simultaneous maximization, in the proposed feedback scheme, of the approximation and regularity (see Section 3.3). Regarding the efficiency of using both types of correlograms, the unidimensional correlogram achieved better approximation errors and more regularity from the first steps, whereas the bidimensional correlogram needed around eight iterations to achieve accurate results. With unidimensional correlograms we needed just two iterations in order to achieve almost the same error than with bidimensional correlograms. This is due to the robustness the unidimensional definition has against shape changes. However, both correlograms are valid for obtaining accurate correspondences with not many iterations. Finally, we see how a registration algorithm such as the one employed by Belongie et al. [4] without any cooperative feedback scheme can lead to poor results when dealing with the types of images we have. We take the same couple displayed in Fig. 11, and only use the landmarks from one type of structure (calcium plaque in this case) for registration. We do so because the descriptor of Belongie is not suitable for different types of structures. Figures 15(a)–(b) show the disposition of the mentioned landmarks in both images. We use Belongie’s registration algorithm with his shape-context descriptor rotated along the tangents of the points in order to achieve orientation invariance. Figure 15(c) shows the final set of correspondences. The points structures are not mapped close to their destination, and the final correspondences are quite irregular. Figure 15(d) shows the transformed image I1 according to these correspondences, and Fig. 15(e) shows in red the edges of this transformed image superposed onto the target image. The warping is very irregular as it is based on irregular correspondences. The


(a)

(b)

(d)

47

(c)

(e)

Figure 15. (a)–(b) Points from the plaque structures of Fig. 11 (a)–(b). Final set of correspondences (c) and final transformation: warped query (d), and edges of warped query superposed on the complementary (e) obtained with Belongies’s registration algorithm.

(a)

(b)

Figure 16. Approximation error (a) and degree of irregularity (b) through the iterations using the shape context matching from [4].

irregularity is due to the high shape difference in both objects, which arises the necessity to enforce step by step some regularity in the correspondences. If this regularity is not enforced, the resulting transformation will map the points without preserving the spatial coherence. Thus, correlograms are good for modelling contextual and global information (as shown with the result in Fig. 13), but only if we strengthen the spatial coherence of the mapping by some feedback algorithm such as the explained in Section 3.3. Finally, Fig. 16 shows the evolution of the approximation error (Fig. 16(a)), and the evolution of the degree of irregularity of the correspondences (Fig. 16(b)). They use 6 iterations in the algorithm. Both the approximation error and irregularity do not decrease throughout the iterations, as the correspondences become more and more irregular.

48


(a)

(b)

(c)

(d)

Figure 17. Registration with mean approximation error.

5.4. Registration Results We see registration results now, using 1646 pairs of homologous images (i.e. from the same category) for computing the statistics. We obtain a mean approximation error of 4.6 pixels, a median error of 2.04 pixels and a standard deviation of 7.6 pixels. The mean distance between two neighbor characteristic points is of 3.1 pixels. Thus, the mapping is 1.5 times the distance between landmarks that are neighbors. Experimentally we have seen that errors below 6 pixels are fairly good. 75% of the alignments have an error below 4.23 pixels. In Fig. 17 we show the alignment with an error of 5 pixels, higher than the mean. Image (a) displays the couple, the left IVUS being the one to be aligned with the right IVUS. Red points are landmarks automatically classified as belonging to plaque structure, whereas blue points are classified as adventitia structure. In cyan we display polygons containing the correct manual classification of plaque structure. Note the high difference in the shapes of the two contours. Images (b)–(c) show the correspondences for the plaque structure, and image (d) shows the global alignment of the two contours: in blue the contour of I2 and in red the mapped contour of I1 . Most of the error in alignment is due to the classification error. Although this error is not high (90.1% of success), the problem is that the points classified incorrectly can spatially concentrate. This makes false structures appear, as shown in Fig. 18. Figure 18(a) shows the automatic classification of landmarks for the couple of images (I1 on the left and I2 on the right), with cyan boxes indicating the correct classification of plaque structure. The classification of the landmarks from I1 is quite correct, but in I2 a false plaque structure is detected. Figure 18(b) shows the computed correspondences. As can be seen, the plaque structure from I1 is matched with the false plaque structure appeared at I2 . 5.5. Retrieval Results For assessing the retrieval efficiency, a database of 100 IVUS images has been used. Two examples of these categories are illustrated in Figs 19 and 20. Figure 19 shows a subset


(a)

49

(b)

Figure 18. Registration with high alignment error due to the detection of false structures.

Figure 19. First example of a subset of images falling into the same category.

of IVUS from a first category. We can see that in all of them the plaque structure has a high degree of embracement around the vessel. Figure 20 shows a subset of IVUS from another category. These IVUS present several plaque structures along the vessel. In this manner, the IVUS images in the database have been classified by a group of physicians into several categories. The categories are chosen attending to clinical properties (see [1]). We extract each image from the database, present it as query, and the system orders the rest of the images from the database in order of similarity to the query. From this ordered list, we take only the first K images (i.e. the K most similar images to

50


Figure 20. Second example of a subset of images falling into the same category.

our query). Two measures of retrieval efficiency are taken. The first one is the estimated number of images we need to retrieve in order to include an image from the same category as the query. We obtained an average of 2.33 images necessary for including one of the same category. For K = 2 retrieved images the mean number of times in which an image from the same category is included is 89.7%. The second measure is the recall vs scope [15]: if query Q has N images from the same category, we compute for Q: E Q (K) =

#|I : rank(I ) ≤ K, category(I ) = category(Q)| N

and average over all the queries presented: E(K) =

1 Q E (K), NQ Q

where NQ is the number of query images presented to the database. We compare the efficiency of the Context Based Plaque Retrieval with the efficiency of the retrieval reported in Huang et al. [15]. They report results using two descriptors: Color Coherent Vectors (CCV) and auto-correlograms, both of them representing types of histograms that include spatial information. In Table 1 we present the E(K) measure for different values of K when using the proposed method versus the Color Coherent Vector-based method, and


51

Table 1. Recall vs scope measure. K

CCV

color auto-correlograms

proposed method

10

0.19

0.38

0.22

30

0.38

0.63

0.50

50

0.38

0.75

0.73

Figure 21. Example of optimal retrieval result for K = 3.

the auto-correlogram. A Color Coherent Vector consists of two histograms, one of noncoherent color pixels of the image, and the other one of color-coherent pixels. A color coherent pixel set is a set of pixels that have a similar color and are spatially contiguous. Auto-correlograms are histograms measuring for some color and some distance the proportion of couples of pixels having this color and lying at this distance of each other. The results with CCV and auto-correlograms are obtained with the database reported in Huang et al. [15]. Huang uses a general database, as CCV and auto-correlograms are not intended to be used in medical domains such as the presented. We only make the comparison for illustration purposes of the goodness of these numbers. As the mean number of images from the same category is 16.56, we show the numbers obtained with CCV and auto-correlogram for experiment on query 4 (see [15]). The efficiency of the proposed approach can be considered to be between the efficiency of CCV and auto-correlograms. In Fig. 21 we see a result of three queries with high performance. In the three cases the first K = 3 retrieved images are all of the same category. In Fig. 22 we see a result of three queries with low performance. Although in this example non of the retrieved images strictly belong to the same category as the query, the similarity of the retrieved images and the queries is quite high. For six of the nine retrieved images, the degree of embracement of the plaque in both the query and the retrieved image is the same. Most

52


Figure 22. Example of worst retrieval result for K = 3.

times the failure of the retrieval system is due to the interpretation of false structures (in the local landmark classification), as can be seen in Fig. 18. Although the correlograms presented are quite robust against non perfect classifications, they can not succeed if an entire spatially coherent region is missclassified.

6. Conclusions In this chapter, we presented a content-based image retrieval method that is able to deal with medical images of bodies with large shape and appearance variability. The proposed method uses a highly discriminant feature space that includes all types of information relevant to retrieve images: local, global and contextual information. This is achieved by generalizing the correlogram descriptor for dealing with different types of structures. The main advantage of the proposed generalization of correlograms is that they are robust against non-exact classifications/segmentations of the different structures inside the image. This allows automatic classifications of the structures. On the other hand, specific information can be easily introduced in this correlogram generalization. This makes the feature space designment suitable for different image domains. Specific information on the other hand is fundamental for dealing with complex domains such as medical image, as general features perform poorly in these domains. Finally, the comparison between images is made invariant to elastic transformations by means of registering the images before their comparison. The registration scheme is based on the point-mapping paradigm, the use of the same generalized correlograms, a thin-plate spline transforma-


53

tion based on few landmarks around discriminant regions and a feedback scheme which strengthens the regularity of the final transformation. Registration with TPS and few landmarks is very efficient as compared to other registration methods such as the use of the Navier-Stokes PDE [3]. There are two main lines of work for improving the presented retrieval system. In the particular IVUS domain, there is still work to do in the characterization of the different types of structures. We have seen that the main cause of error in the system is the wrong classification of entire regions in the image. In this sense, we could explore improved local descriptors. The inclusion of information along the longitudinal axis of the arteries is very important, as the physicians use this information for identifying the different types of structures. In general, considering the problem of medical image retrieval, we want to explore fast indexing techniques of the generalized correlograms by use of fast spatial access methods.

References [1] J. Amores and P. Radeva, Elastic matching retrieval in medical images using contextual information, Tech. report, CVC, September 2003. [2] Nur Arad., Image warp design based on variational principles, Ph.D. thesis, Tel-Aviv University, Tel Aviv, 1995. [3] R. Bajcsy and S. Kovacic., Multiresolution elastic matching, Computer Vision, Graphics and Image Processing (1989), no. 46, 1–21. [4] S. Belongie, J. Malik, and J. Puzicha., Shape matching and object recognition using shape contexts., Tech. Report UCB//CSD-00-1128, UC, Berkeley, 2001. [5] Marco Bressan and J. Vitria, Nonparametric discriminant analysis and nearest neighbor classification, Pattern Recognition Letters 24 (2003), no. 15, 2743–2749. [6] L. Brown., A survey of image registration techniques, ACM Computing Surveys 24 (1992), no. 4, 325– 376. [7] C. Carson, M. Thomas, S. Belongie, J.M. Hellerstein, and J. Malik, Blobworld: A system for regionbased image indexing and retrieval, Third Int. Conf. on Visual Information Systems (Springer-Verlag, ed.), LNCS, 1614, 1999, pp. 509–516. [8] S.K. Chang, C.W. Yan, D.C. Dimitroff, and T. Arndt, An intelligent image database system, IEEE Transactions on Software Engineering 14 (1988), no. 5. [9] Y. Chen and J. Wang, A region-based fuzzy feature matching approach to content-based image retrieval, IEEE Transactions on Pattern Analysis and Machine Intelligence 24 (2002), no. 9, 1252–1267. [10] J. Dahmen, T. Theiner, D. Keysers, H. Ney, T. Lehmann, and B. Wein, Classification of radiographs in the ’image retrieval in medical applications system (irma), Procs 6 th Int. RIAO Canf on Content-Based Multimedia Information Access (Paris), 2000, pp. 551–566. [11] J.G. Dy, C. E. Brodley, A. Kak, L.S. Broderick, and A.M. Aisen, Unsupervised feature selection applied to content-based retrieval of lung images, IEEE TPAMI 25 (2003), no. 3. [12] Boston Scientific Europe. (ed.), Beyond angiography. intravascular ultrasound: State of the art, vol. 1, XX Congress of the ESC, August 1998. [13] M. Flickner, H. Sawhney, W. Niblack, J. Ashley, Q. Huang, B. Dom, M. Gorkani, J. Hafner, D. Lee, D. Petkovic, D. Steele, and P. Yanker, Query by image and video content: The qbic system, IEEE Computer (1995), 23–32. [14] T.-Y. Hou, P. Liu, A. Hsu, and M.-Y. Chiu., Medical image retrieval by spatial features, IEEE Int. Conf. on Systems, Man and Cybernetics., vol. 2, 1992, pp. 1364 –1369. [15] J. Huang, S. Kumar, M. Mitra, W. Zhu, and R. Zabih, Image indexing using color correlograms, Proc. CVPR., 1997, pp. 762–768. [16] F. Korn, N. Sidiropoulos, C. Faloutsos, E. Siegel, and Z. Protopapas, Fast nearest neighbor search in medical image databases, Pr. of the 22nd VLDB Conference, 1996, pp. 215–226. [17] S.Y. Lee and F.J. Hsu, 2d c-string: A new spatial knowledge representation for image database systems, Pattern Recognition 23 (1990), no. 10, 1077–1087.

54


[18] Lifeng Liu and Stan Sclaroff, Medical image segmentation and retrieval via deformable model, International Conference on Image Processing (Thessaloniki , Greece), vol. 3, October 2001, pp. 1071–1074. [19] Yanxi Liu and Frank Dellaert, Classification driven medical image retrieval, Proc. of the Image Understanding Workshop, 1998. [20] Yanxi Liu, Frank Dellaert, William E. Rothfus, Andrew Moore, Jeff Schneider, and Takeo Kanade, Classification-driven pathological neuroimage retrieval using statistical asymmetry measures, Proceedings of the 2001 Medical Imaging Computing and Computer Assisted Intervention Conference (MICCAI ’01) (Utrecht, The Netherlands), October 2001. [21] G.S. Mintz, J.J. Popma, C.J. Ditrano, J. Mackenzie, and L.F. Satler, Intravascular ultrasound vs. quantitative coronary angiography: A statistical comparison of 538 consecutive target lesions (abstract), Circ. 88 (1993), 1–411. [22] C. Papadimitriou and K. Stieglitz, Combinatorial optimization: Algorithms and complexity, 1982. [23] Paredes, D. Keysers, T.M. Lehmann, B.B. Wein, H. Ney, and E. Vidal., Classification of medical images using local representations, Bildverarbeitung fur die Medizin, 2002, pp. 171–174. [24] A. Pentland, R.W. Picard, and S. Sclaroff., Photobook: Tools for content-based manipulation of image databases, Storage and Retrieval for Image and Video Databases (SPIE, ed.), 1994, pp. 34–47. [25] E.G.M. Petrakis and C. Faloutsos, Similarity searching in medical image databases, IEEE Transactions on Knowledge and Data Engineering 9 (1997), no. 3. [26] M.J.D. Powell, A thin plate spline method for mapping curves into curves in two dimensions, Computational Techniques and Applications (CTAC95) (Melbourne, Australia), 1995. [27] G.P. Robinson, H.D. Tagare, J.S. Duncan, and C.C. Jaffe, Medical image collection indexing: Shapebased retrieval using kd-trees, Computerized Medical Imaging and Graphics 20 (1996), no. 4, 209–217. [28] C. Schmid and R. Mohr, Local grayvalue invariants for image retrieval, IEEE TPAMI 19 (1997), no. 5. [29] C.R. Shyu, C.E. Brodley, A.C. Kak, A. Kosaka, A.M. Aisen, and L.S. Broderick, Assert - a physician-inthe-loop content-based retrieval system for hrct image databases, Computer Vision Image Understanding (1999), no. 75, 111–132. [30] M.J. Swain and D.H Ballard, Colour indexing, International Journal of Computer Vision 7 (1991), no. 1, 11–32. [31] James Ze Wang, Jia Li, and Gio Wiederhold, SIMPLIcity: Semantics-sensitive integrated matching for picture LIbraries, IEEE Transactions on Pattern Analysis and Machine Intelligence 23 (2001), no. 9, 947–963.


55

MRI Plaque Tissue Characterization and Assessment of Plaque Stability Chun YUAN a , Thomas S. HATSUKAMI b and Jianming CAI a a University of Washington, Seattle, WA, USA b Department of Surgery and VA Puget Sound Health Care System, Seattle, WA, USA Abstract. This chapter reviews current MRI techniques to differentiate stable versus high risk atherosclerosis and discusses the development of non-invasive MR imaging techniques to characterize atherosclerotic plaques. Tissue specific MR signal features will be described according to histo-pathological evaluation standards and comprehensive imaging protocol for the identification of different lesion types will be introduced. Keywords. Atherosclerosis, magnetic resonance, black blood, white blood, wall morphology, intraplaque hemorrhage, histology, quantitation

1. Introduction Stroke is a leading cause of death (4.4 million per year) and disability (5000 per million persons) worldwide. Due to the aging of the population, the rate of stroke is projected to increase and become an even greater healthcare concern. In the United States alone, the rate of stroke-related death is predicted to outpace population growth and may double over the next 30 years [1]. Better methods to detect atherosclerotic disease would considerably lower stroke-related mortality rates, improve the quality of life of stroke survivors, and reduce health care costs. Carotid atherosclerosis is a major contributor to the etiology of ischemic stroke. Autopsy studies have provided ample pathological evidence of athero-embolic material in the small arteries that supply infarcted tissue in the brain [2]. Several large randomized clinical trials have shown that the removal of high-grade carotid plaques (carotid endarterectomy) significantly reduces incidence of recurring stroke compared to non-operatively treated individuals with similar degrees of carotid stenosis [3–6]. Secondary analysis of data from the North American Symptomatic Carotid Endarterectomy Trial (NASCET) has shown that the features of the carotid atherosclerotic lesion, such as plaque surface characteristics, are associated with a dramatically increased risk for stroke. For example, angiographic identification of plaque ulceration increased the 2-year risk for ipsilateral stroke to 73% amongst those with 95% stenosis, compared to 21% of patients with non-ulcerated 95% carotid stenosis [7]. Although the randomized clinical carotid surgery trials have provided new insights into the role of carotid endarterectomy in the management of stroke patients, a better

56

C. Yuan et al. / MRI Plaque Tissue Characterization and Assessment of Plaque Stability

understanding of the mechanisms of carotid-related stroke is needed. A number of studies suggest that in addition to the plaque-lumen surface characteristics, the composition and morphology of the plaque itself are important determinants of risk of thromboembolic complications. Based on histopathological findings from excised surgical or autopsy specimens, the high-risk, vulnerable lesion has been typically described as containing a large necrotic core or intraplaque hemorrhage that is separated from the lumen by a thin fibrous cap with inflammatory cellular infiltration [8–11].

2. Need for Atherosclerosis Imaging Currently, histopathological study of atherosclerotic plaque specimens obtained at autopsy is the primary method used to examine plaque characteristics. A shortcoming of this approach is that its assessment of such characteristics is limited to a single timepoint. Furthermore, research access to a wide spectrum of atherosclerotic lesions is limited. In order to better understand the fundamental mechanisms involved in the development and progression of high-risk plaque characteristics, serial in vivo assessment of the lesion is needed. To meet this need, the development of accurate, reproducible and preferably non-invasive imaging methods are required. High-resolution atherosclerosis imaging is also needed to develop novel pharmacological therapy. Recent developments in animal models of atherosclerosis are promising, but to date, they are unable to mimic the advanced lesions of human atherosclerosis. As a result, many pharmacological interventions that appeared promising in animal models were not effective when applied in human trials, resulting in significant waste of health care resources. High-resolution atherosclerosis imaging is therefore needed to provide direct, serial, and in vivo assessment of how human atherosclerosis responds to therapeutic interventions. Given its superficial location, relatively large size and immobility, and the presence of the full spectrum of American Heart Association (AHA) lesion types [12], the human carotid artery is better suited for serial imaging studies than other arterial sites. More importantly, the carotid artery is the only vascular bed where intact surgical specimens are readily available for histological verification of findings from novel in vivo imaging techniques (Fig. 1). Magnetic resonance imaging has demonstrated great promise as an accurate, reproducible method for characterizing human carotid atherosclerosis in vivo. MRI has several inherent advantages compared to other imaging modalities: 1) it is non-invasive; 2) through advances in hardware and image acquisition protocols, sub-millimeter resolution is achievable; 3) it has superior capability to distinguish tissue types based on their chemical composition; and 4) advances in targeted MRI contrast agents show promise for superior identification of specific molecular targets. In the sections that follow, the technical aspects of high-resolution MR imaging will be reviewed and validated.

3. Technical Aspects of Plaque Imaging Clinical carotid plaque imaging techniques that use a 1.5T clinical scanner set to the spatial resolution of 1 × 1 mm2 are unable to accurately detect the small volumes of


57

Figure 1. The complexity of the carotid atherosclerotic plaque is well illustrated in this section taken from the common carotid portion of an excised endarterectomy specimen.

the major plaque components present in carotid lesions. Recent research shows that the evaluation of histologically processed endarterectomy specimens reveal the mean values for the volume of individual plaque components (lipid core, fibrous intimal tissue, and calcification) that range from 0.3 mm3 and up [13]. While submillimeter voxel sizes can be achieved with whole body 1.5T scanners, it is necessary to use a phased-array surface coil to generate an acceptable signal-to-noise ratio (SNR) [14] for the high-resolution images required for carotid plaque characterization. Since high spatial resolution imaging sequences may also result in relatively long scan times (several seconds to a few minutes), hardware and software techniques that reduce the effects of motion and flow artifacts are key to improving imaging quality. The following sections describe the hardware considerations and the imaging sequences that are valuable for producing high quality diagnostic multi-contrast MR images of the carotid arteries. 3.1. Hardware Considerations The carotid artery is a superficial structure with lengths greater than their distance from the skin surface. Its configuration makes possible the use of phased-array (PA) surface coils to simultaneously collect data from both its right and left sides. Accordingly, a dedicated PA coil assembly [15] with the dimensions of 6.4 × 10.8 cm was constructed to image carotid plaques. To make possible the imaging of both carotid arteries during the exam, the assemblage consisted of two separate sets of coils. Each coil is made of a soft flexible material that can be comfortably fitted and secured around the patient’s neck (Fig. 2). This coil assembly makes it possible to obtain data from a carotid longitudinal

58


Figure 2. A picture of the carotid phased array coil. Notice the shape of the coil surface is designed to touch the human skin. The coil is flexible for different size of the human neck.

coverage of up to 5 centimeters. Performance studies of these PA coils have shown a 37% improvement in the SNR when compared to commercially available 3-inch surface coils. This enhancement of the signal-to-noise ratio (SNR) enables the acquisition of diagnostic images of the common, internal and external carotid arteries with an average voxel size of 0.25 × 0.25 × 2.0 mm3 (0.25 mL). In addition to the carotid phased-array assembly, a custom designed headholder was constructed using vacuum formed PVC plastic. The headholder provides support for the occiput and neck, which not only improves patient comfort, but also makes possible repeatable scan positioning and reduces patient movement. 3.2. Black and Bright Blood Pulse Sequences T1 and T2 weighted images were the first to be used to identify individual plaque components [16,17]. Recent research, however, has shown that the use of multiple contrast weightings can significantly improve the precision and accuracy of MRI characterizations of plaque morphology [18–20]. To improve the quality of MRI depictions of plaque morphology, these studies recommend using a combination of spin echo (SE) based T1, T2 and proton density weighted images and gradient recalled echo (GRE) T2* images [21]. The SE techniques are primarily used for studying plaque morphology and tissue characterization — especially lipid, hemorrhage and fibrous tissues. The GRE techniques (similar to those of the TOF) are used for studying lumen-plaque interface and plaque morphology. In other words, the pulse sequences designed for vascular imaging are either, depending on the signal of the blood flow relative to the surrounding soft tis-


59

sues, black blood or bright blood techniques. These two types of sequences offer specific advantages for carotid imaging. Black blood techniques are MR imaging techniques that suppress the signal obtained from blood flow [22]. This technique is ideal for plaque imaging because the conspicuity of the vessel wall is increased when it is next to a hypointense lumen and the imaging echo (TE), and repetition (TR) times can be varied to optimize visualization of specific plaque components. The major disadvantages of black blood techniques include relatively long scan times, and the fact that these sequences are based on 2D data acquisitions requiring slice thickness between 2 and 5 mm. Common flow suppression techniques employed with black blood imaging include the following: the use of pre-saturation radio frequency bands applied along the direction of the arterial blood flow with a spin-echo sequence and a double inversion recovery (DIR) sequence [23]. When pre-saturation techniques are used — which are less effective than DIR with slowly flowing blood — the complex flow in the carotid bulb [24] often results in artifacts created by unsuppressed flow. Artifacts may be misinterpreted as representing a signal from a diseased vessel wall, which leads to overestimations of the size of the atherosclerotic lesion (Fig. 3) [25]. On the other hand, DIR sequences tend to provide excellent flow suppression as shown in T1 weighted images [26]. These images typically provide the most accurate quantitative measurements of disease burden and are thus used to identify necrotic cores in vivo [21]. Bright blood techniques refer to the gradient echo based imaging sequences that are typically used to acquire MR angiograms. These sequences, such as GRE and SPGR, enhance the signal of flowing blood, and thus the lumen appears as hyperintense in relation to the adjacent vessel wall. Compared to SE sequences, bright blood techniques can produce images with shorter TE and TR times. The lack of a spin-echo in these sequences creates T2* sensitive tissue signals that appear to improve the visualization of the intimal calcifications and the fibrous cap, which is typically a dense structured layer of collagen [26,27]. Faster scanning also allows for the acquisition of high-resolution 3D data sets that will improve plaque characterization [28]. Based on the work of Hatsukami et al., we currently incorporate a 3D-TOF sequence to provide GRE weighted axial images to characterize carotid plaque. These images have been useful for evaluating the in vivo state of the fibrous cap [26] and for detecting large intraplaque hemorrhages [21]. 3.3. Imaging Protocol of Multiple Contrast Weighting Based on extensive testing with normal volunteers and endarterectomy patients, a standardized, multicontrast imaging protocol was developed to assess the in vivo morphology of carotid plaques. This protocol uses black and bright blood techniques to acquire high spatial resolution axial images of the carotid arteries in each subject and provides an oblique view of the carotid artery, which larifies the location of the carotid bifurcation process and the development of plaque distribution. It also uses the bifurcation area as an internal landmark to reproducibly prescribe slice locations for serial studies, and limits the total exam time to an average of 40 minutes. Currently, three axial imaging sequences (3D-TOF, T1W, PDW and T2W) are performed to generate four different contrast weightings at each slice location (Fig. 4). Applying a zero-filled Fourier Transform [29] to all imaging sequences, a voxel size of 0.25 × 0.25 × 2.0 mm3 was achieved for the black and bright blood sequences. Though the imaging protocol varies depending

60


Figure 3. Comparison between the routine Spin Echo (SE) T1WI and Double IR T1WI. Notice the good luminal boundary delineation achieved from using the DIR technique. The long white arrow points to the bifurcation of left carotid artery. The * sign indicates the location of lumen. The flow artifact is significantly suppressed by DIR T1WI (short white arrow).

on patient body habitus, the typical set of parameters for the three axial sequences that were used are summarized in Table 1. Chemical selective fat saturation is also used to reduce, in all sequences, the subcutaneous tissue signal [30,31]. Cardiac Gating was found to reduce the flow and motion artifacts and is incorporated in the long TE and TR sequences. 3.4. Histology as the Gold Standard Carotid endarterectomy of the atherosclerotic plaque creates a unique opportunity to verify by histological examination tissues visualized by in vivo MRI. Specimens are surgi-


61

Figure 4. A set of images taken from a patient with severe carotid stenosis. The right hand figure is an oblique FSE view of the right carotid artery. PDWI and T2WI images are cross sectional images at the internal carotid artery. TOF image of the same location shows bright lumen (short white arrow). T1WI is a DIR FSE image of the same location. The lesion is eccentric (long white arrow) and caused a severe lumen stenosis.

Table 1. Image parameters. TOF

T1W∗

IW

T2W

TR (ms) TE (ms) Thickness (mm) Matrix Excitations ETL∗∗∗

23 3.8 2 256 × 256∗∗ 2

800 9 2 256 × 256∗∗ 2 8

Gated 20 2 256 × 256∗∗ 2 8

Gated 40 2 256 × 256∗∗ 2 8

Scan time (min) FOV (cm) Other Other

1.5-4 13 Fat suppression

6-7 13 Fat suppression

5-7 13 Fat suppression Spatial flow

5-7 13 Fat suppression Spatial flow

saturation

saturation

∗ T1W: Double inversion recovery with a TI of 650 ms was used. ∗∗ Zero-filled Fourier transform was used to create images of 512 × 512 matrix. ∗∗∗ ETL: echo traim length for FSE sequences.

cally excised using a technique that allows the specimen to remain intact. Fixation is accomplished with the standard 10% neutral buffered formalin. 10% formic acid is used to dissolve calcifications that would interfere with sectioning. The specimen is processed, paraffin embedded and sectioned en-bloc. 10 μ sections are collected at every millimeter of the common carotid to the bulb and at 0.5 mm through the bulb and internal carotid.

62


Figure 5. Carotid endarterectomy specimens are histologically processed en-block and serially sectioned producing 10 to 20 cross sections per centimeter. The varied composition of the carotid atherosclerotic plaque is well illustrated in this specimen.

The hematoxylin and eosin stain, special stains and selected antibodies are used to identify the morphological components of the plaque. Precision in matching the numerous histology sections to the MR images is crucial to the understanding of the capability and accuracy of MRI. Lumen size, shape, distance from the bifurcation and pieces of calcification are all used to attain the correct match. The histology sections are photographed and digitized. Plaque components are outlined and quantitative data is extracted that can be matched to similar data collected from outlined MR images. The status of the fibrous cap, size of the necrotic core, amount and position of calcifications and intraplaque hemorrhage all play an important role in the stability of the atherosclerotic plaque. These features are easily visualized with the light microscope making histology the gold standard used by MR imaging in the development of sequences and image analysis tools (Fig. 5). 4. MR Characterization of Human Carotid Plaque Morphology The vessel luminal diameter and area as measured by angiographic techniques is clinically used to describe plaque morphology. As MRI can visualize the luminal and the outer wall boundary as a series of plaque, size related parameters can offer a more comprehensive measure of plaque morphology in terms of plaque volume, thickness, and 3D distribution. MRI is therefore uniquely positioned to study in vivo the effects of compensatory enlargement of the artery.


63

Figure 6. Cross-sectional image of carotid artery on in vivo (A) T1-weighted MRI. On Figure B, the area of low signal (short black arrow) represents a region of calcification, and the area of high signal (long black arrow) represents a region of recent hemorrhage. Outline of lumen (red) and the internal carotid artery outer wall boundary (green circle) are showed, and the areas can be measured. On Figure C, the Maximum Wall Thickness (long yellow line) and Minimum Wall Thickness (short yellow line) are showed.

4.1. Plaque Area/Volume Measurement Unlike lumen stenosis, plaque volume is considered a direct measure of the size and severity of atherosclerotic disease. Because MR is able to identify the adventitia boundary in transverse images of the vessel wall, MR imaging could provide a means to measure the total volume of the diseased vessel wall and to accurately determine the composition of plaque burden (Fig. 6) [32]. The high quality tissue contrast of the areas between the diseased portions of the vessel, lumen, and adventitia have made possible experiments that evaluate the quantitative capabilities of in vivo MR to evaluate the accuracy of such measurements and to corre-

64


late vessel wall measurements to those used in conventional clinical settings (stenosis). One such study compared the images of cross-sectional areas of carotid plaques to those measured preoperatively with similar area measurements made of the excised endarterectomy specimens that were imaged ex vivo. A Bland-Altman analysis performed on the paired in vivo and ex vivo measurements of the same vessel segments demonstrated a strong correlation between the two values [32]. These results confirmed the quantitative capabilities of MR for measuring total plaque volume and disease burden. 4.2. Accuracy of MRI Based Vessel Wall Morphology Measurement In a follow up study [33] by the same group, the accuracy and precision of the measurements of carotid plaque burden and lumen narrowing were determined. The study configurations was similar to the one used in [32]. In brief, a group of 37 patients who underwent CEA were scanned before surgery using a black blood MRI technique and an ex vivo MRI of the excised specimen was conducted immediately after surgery using the same technique. Three different plaque measurements were obtained and compared between paired in vivo and ex vivo MR images: maximum wall area (MWA), minimum lumen area (mLA), and wall volume (WV). MWA and WV are measures of plaque burden, while mLA is a measure of lumen narrowing. The in vivo and ex vivo measurements highly correlated to each other (the correlation coefficients for in vivo/ex vivo WV, MWA, and mLA were 0.92, 0.91, 0.90 respectively). This study therefore shows that in vivo black blood MRI can be used to accurately estimate the morphology of the plaque. The minimum lumen area measurement (lumen narrowing) and the MWA or the WV were not highly correlated (the correlation coefficients between mLA and MWA or WV < 0.3). This finding suggests that outer wall boundary is also actively involved in the atherosclerotic disease process and that MRI is a useful tool to study the relationship of these different aspects of morphological changes. 4.3. Reproducibility Since MRI is able to non-invasively identify both the luminal and outer wall boundaries, it can also be used to study the morphological changes of atherosclerosis. A first step in such a direction is to test the inter-scan reproducibility of quantitative measures of plaque burden by MRI, which was assessed by two recent experiments [34,35]. By using the carotid bifurcation as an internal landmark, these studies showed that image slices of the same vessel segment can be reproducibly obtained at different exam times. One of these studies analyzed the precision of quantitative measurements for both the lumen and vessel wall areas of human carotid arteries [35]. Based on data obtained from independent MR scans conducted on 8 patients in the period of two weeks, the error of lumen area measurement was 6.2%, 9.2% and 9.7% for T1W, IW, and T2W images, respectively. The estimation of error is based on standard analyses of variance calculations of the mean and the standard deviation of the area measurements from pooled locations. Wall area measurement error was 10.8%, 10.9%, and 12% for the three contrast weightings. Errors in wall volume measurement ranged from 4–6% across different contrast methods. The vessel wall volume therefore can in fact be measured accurately. Among the factors that impact MRI area measurement, image quality is the most important.


65

Table 2. Contrast of the main components of atherosclerotic plaque. MR Contrast Weighting T1W IW

Plaque Component

TOF

Recent Hemorrhage Lipid Rich Necrotic Core Intimal Calcifications

++ +/− −−

++ ++ −−

Variable Variable −−

Variable Variable −−

+/− to −−

+/−

+

Variable

Fibrouis Tissue

T2W

∗ Tissue contrast is relative to the signals from sternocleidomastoid muscle.

4.4. Image Processing Techniques Needed for Plaque Morphology Evaluation Please refer to “MRI Based Quantitative Analysis System for Atherosclerotic Tissue Evaluation: CADCASDE”.

5. MR Characterization of Human Carotid Tissue Composition Histopathological studies indicate that the typical morphology of the vulnerable plaque consists of a large necrotic core that is separated from the lumen by an unstable fibrous cap that can be thin, ulcerated, fissured, or infiltrated by macrophages. The initial appeal of MR techniques for imaging atherosclerosis was its theoretical ability to image the lipids in the necrotic core. Early attempts to configure MRI to image atherosclerosis used T1W and chemical selective techniques to focus on the detection of lipid signals using T1W [36–39]. The predominant lipids of atherosclerotic lesions are cholesterol and cholesterol esters [40] rather than triclycerides. Since these lipids have short T2 components, initial attempts at plaque imaging produced limited success. It was the inclusion of either T1W and T2W sequences, or proton density weighted (T2W, T1W, or PDW) with ultra-short echo times (TE) that improved the specificity of MR imaging techniques and allowed for the differentiation of necrotic cores from fibrous regions in ex vivo plaque specimens [16,41]. Follow up research added contrast weightings such as 3D-Time-of-Flight (TOF) [26] to magnetization transfer (MT) [42,43] and showed that diffusion sensitive [18,44,45] sequences facilitate atherosclerotic tissue characterization. These studies suggest that the major plaque components — lipid rich necrotic core, calcium deposits, fibrous connective tissue, and intraplaque hemorrhage — can be identified by their signal characteristics in T1, T2, and proton density weighted images (Table 2) [17–19]. The six studies described below is informed by the research reviewed above and establish the accuracy of MRI in plaque tissue carotization. These studies used the same imaging protocol to produce four different contrast weightings — 3D-TOF, T1 weighted double-inversion recovery (DIR) SE sequence, and the SD and T2 weighted images from a shared-echo SE technique — at each imaged location of the carotid arteries of eighteen endarterectomy patients. All scans were conducted on a GE 1.5T Signa scanner using phased array coils. 5.1. In Vivo Accuracy of MR for Detecting Unstable Fibrous Caps Hatsukami et al. were the first to use a 3D-TOF bright-blood imaging technique (multiple overlapping thin slab angiography or MOTSA [46]) to identify unstable fibrous caps

66


Figure 7. Example of plaque with fibrous cap rupture on gross section, histology, and MRI. Figures A–C are three contiguous axial 3D TOF images of a diseased right CCA, D is a photo of the gross section at corresponding location, E is a low power photomicrography of a histology section with Masson’s trichrome stain (10×), and F is the same MR image as C. On the gross and histological sections, there is an area of cap rupture (arrow 1) across from a region where the fibrous cap is thick (arrow 3). The cap rupture site corresponds to a region where the hypointense band is absent, and a hyperinternse region is seen adjacent to the lumen on MRI. The hyperintense regioin is a region in the plaque core on MRI that corresponds to regions or recent intraplaque hemorrhage on gross and histological cross sections (arrow 2). The fibrous cap is made of a dense layer of collagen which appeared hypointense on the 3D TOF images.

in the atherosclerotic in vivo human carotid arteries [26]. The images were acquired using a spoiled gradient echo (SPGR) sequence that was set to a low flip angle to better visualize both flowing blood and soft tissue. The tissue contrast of these images were weighed more heavily on the T2* side than on the shorter TE side. After careful review of twenty-two preoperatively imaged endarterectomy patients, the authors found that the histological state of the fibrous cap correlates well (Kappa value of 0.83) with the images of the hypointense juxtaluminal bands. Hypothesizing that the layered organization of the fibrous cap was responsible for the relative signal loss in the MR images, they were able to prospectively characterize the in vivo state of the fibrous cap as intact and thick, intact and thin, or ruptured (Fig. 7). These findings provided the basis of a research project that evaluated the ability of multi-contrast MR to characterize the in vivo state of the fibrous cap [21]. In agreement with the work of Hatsukami et al., the results demonstrated a strong correlation between the MR image findings and the histological state of the fibrous cap (Kappa value of 0.71). The authors were thus able to show that the availability of the three SE contrast weightings facilitated image interpretation in 17 of the 91 image locations The larger sample size made it possible to report test performance statistics (sensitivity of 81%, specificity of 90%) for noninvasive identification of unstable fibrous caps in vivo.


67

5.2. In Vivo Accuracy of MR for Identifying Soft Necrotic Cores An important morphological characteristic of the vulnerable plaque is the presence of a large soft core that is comprised of a cellular lipid rich region or hemorrhage [47]. This study evaluated the in vivo accuracy of MR to detect the soft cores in human carotid plaques [21]. Based on the plaque tissue contrast features described in Table 2, lipid-rich, necrotic cores were identified by areas of hyper-intensive signal intensity in the T1W images that were of iso-intensity in the corresponding 3D-TOF images, and were of variable intensity on PDW and T2W. Recent intraplaque hemorrhages were similarly identified by their multi-contrast appearance (high signal on TOF, high on T1W, and of variable intensity on PDW and T2W). The results revealed that the MR findings agreed with either the histologic presence of a necrotic core or the recent intraplaque hemorrhage, with a sensitivity of 85%, specificity of 92%, and a calculated Kappa value of 0.69. The variation of signals from intraplaque hemorrhage of different ages was also observed but not yet documented. 5.3. Intraplaque Hemorrhage Aging and Its Corresponding MRI Signal Variations In clinical settings, knowledge of the age of intraplaque hemorrhage can give insight into the history and current condition of the biologically active plaque. Signal variations of brain hemorrhage is a well studied and reasonably understood problem [48]. This study used high-resolution MRI to establish reliable criteria for detecting hemorrhagic incidents in carotid atherosclerotic plaques and identify their evolutionary stage. Two readers — both unaware of histological data on the patients — reviewed the MR images and grouped hemorrhage into three categories (early, recent and old) using a modified cerebral hemorrhage criteria. Hemorrhage was histologically identified and staged in 145/189 (77%) of carotid artery plaque quadrants. MRI was able to detect intraplaque hemorrhage with high sensitivity (90%) but moderate specificity (74%). There was moderate to good agreement between MRI and histology classifying stages (Cohen’s κ = 0.7, 95% CI: 0.5–0.8 for reviewer 1 and 0.4, 95% CI: 0.2–0.6 for reviewer 2), and moderate agreement between the two MRI readers (κ = 0.4, 95% CI: 0.3–0.6). 5.4. MRI Based Atherosclerotic Lesion Type Definition Since in vivo MRI is able to identify the composition of the carotid atherosclerotic plaque, it follows that MRI may also be able to re-define lesion types. AHA definitions of lesion types aims to provide a unified method to stage lesions [49]. This study [50] used the AHA lesion type classification system to determine the accuracy of in vivo high-resolution multi-contrast MR imaging of the human carotid atherosclerotic plaque. Using a 1.5T scanner, carotid endarterectomies were performed on 60 atherosclerotic patients (mean age 70, male 54). During the procedure, carotid plaques were removed intact and processed for histological examination. In order to decrease the dependency of MRI information on adjacent image locations, images were selected at 4mm distances. AHA lesion type classification was slightly modified for MRI: Type I–II—near normal wall thickness, no calcification; Type III—slightly diffuse thickening or slightly eccentric thickening with no calcification; Type IV–V—plaque with a lipid or necrotic

68


Figure 8. Example of type IV–V lesion in internal carotid artery (lipid-rich necrotic core was detected by histology). On multicontrast-weighted MR images, lipid-rich necrotic core (arrow) had iso signal intensity (SI) on both T1WI and TOF images, but iso-SI to slightly high SI on PDWI and T2WI images. Lumen is moderately stenosed. * indicates lumen. Original magnification ×10. (This Figure is originally published: Cai J.M., et al. Classification of human carotid atherosclerotic lesions with in vivo multicontrast magnetic resonance imaging. Circulation. 2002; 106(11):1368–73).

core surrounded by fibrous tissue with possible calcification; Type VI—complex plaque with possible surface defect, hemorrhage, or thrombus; Type VII—calcified plaque; Type VIII—fibrotic plaque with no or very small lipid core and possible small calcifications. Both MR images and histological sections were independently reviewed and categorized and then compared. Sensitivity and specificity of different lesion types were calculated. Cohen’s Kappa (κ) and weighted Kappa value were computed to quantify the agreement between the MRI findings and histology. Overall, the classification system obtained by MR imaging and the AHA classifications showed good agreement, with κ (95%CI) = 0.74 (0.67 to 0.82) and weighted κ = 0.79. Compared to the histological gold standard, the sensitivity and specificity of MR Imaging classification were: type I–II, 67% and 100%; type III, 81%and 98%; type IV–V, 84% and 90%; type VI, 82% and 91%; type VII, 80% and 94%; type VIII, 56% and 100%. This study thus showed that in vivo high-resolution multi-contrast MRI is able to classify intermediate to advanced atherosclerotic lesions in the human carotid artery and to distinguish between advanced lesions from early and intermediate atherosclerotic plaque (Figs 8 and 9). 5.5. Overall Accuracy of MRI Quantitative Tissue Detection MRI plaque imaging techniques can be used for clinical plaque risk assessment and in clinical trials to evaluate the atherosclerotic lesion progression and regression. Currently, the endpoints used in clinical trials to study plaque progression and/or regression are based on wall thickness measurements. MRI could provide additional information in clinical trials regarding plaque composition. A baseline study to evaluate the ability of


69

Figure 9. Example of type VI lesion just distal to carotid bifurcation (acute to subacute mixed hemorrhages were detected by histology). On multicontrast-weighted MR images, acute and subacute mixed hemorrhage had high SI on both TOF and T1WI images, iso-SI to slightly high SI on PDWI and T2WI images (arrow). * indicates lumen. Original magnification ×10. (This Figure is originally published: Cai JM, et al. Classification of human carotid atherosclerotic lesions with in vivo multicontrast magnetic resonance imaging. Circulation. 2002; 106(11):1368–73).

MRI to not only identify the presence of carotid plaque constituents, but also to measure the quantity of composition material in vivo, compared to histology. In this study, twenty-two randomly selected patients with at least average image quality, scheduled for carotid endarterectomy were imaged with a 1.5T scanner after informed consent. Area measurements of lumen, wall area, lipid rich/necrotic core, hemorrhage, calcium, and loose matrix were compared between 144 in vivo MR images and matched histology sections. These measurements were performed using a customdesigned imaging analysis tool (QVAS). Sensitivity (S) and specificity (SP) is given for each tissue group at the 144 locations. The mean area of each component was calculated per artery and the Pearson correlation (R) coefficient was used to correlate the MRI and histology measurements. The matched in vivo MR images and histology slices showed strong and highly significant correlation for lumen area (R = 0.84), wall area (R = 0.80), size of the lipid rich/necrotic core (R = 0.76; S = 93%, SP = 62%), hemorrhage (R = 0.75; S = 90%, SP = 72%) and calcification (R = 0.78; S = 81%; SP = 89%). Agreement for loose matrix was moderate (R = 0.60; S = 62%; SP = 71%). In summary, MRI can be used to characterize and measure the components of the carotid atherosclerotic plaque. 5.6. Plaque Neovasculature and Inflammation Please refer to “Imaging of Plaque Cellular Activity with Contrast Enhanced MRI”.

70


6. Plaque Features Visualized by MRI and Their Association with Patient’s Neurological Symptoms Histological studies have revealed some of the salient features of vulnerable plaques. These studies have pointed to the importance of the layer of collagen-rich fibrous tissue that separates vessel lumen from the plaque bulk and the presence of necrotic, calcific, and or hemorrhagic tissues near this layer. In a retrospective case-control study [51], the appearance of the fibrous cap on MRI (intact thick, intact thin or ruptured) was shown to be highly associated with recent ischemic neurological symptoms. In this study, 53 consecutive patients (mean age 71, 49 male) scheduled for carotid endarterectomy were recruited after obtaining informed consent. 28 subjects had a recent history of TIA or stroke on the side appropriate to the index carotid lesion, and 25 were asymptomatic. Pre-operative carotid MRI was performed in a 1.5T GE Signa that generated T1-, PD-, T2-weighted, and 3D time-of-flight images. Using previously reported MRI criteria, the fibrous cap was categorized as intact-thick, intact-thin, or ruptured for each carotid plaque by blind review. There was a statistically significant trend showing a higher percentage of symptomatic patients for ruptured caps (70%) compared to a thick cap (9%) (p = 0.001 Mann-Whitney test for cap status vs. symptoms). Compared to patients with thick fibrous caps, patients with thin fibrous caps were 10 times more likely to have had a recent TIA or stroke (95% CI = 1.0, 104), and those with ruptured fibrous caps were 23 times more likely to have had recent ischemic neurological symptoms (95% CI = 3, 210). This study showed that MRI identification of a ruptured fibrous cap is highly associated with a recent history of TIA or stroke.

7. Practical Applications Advances in vascular biology have led to the introduction of new medical therapies that require clinical trials to evaluate their efficacy. Noninvasive characterization of plaque morphology provides biologic endpoints for such therapeutic trials. The establishment of accurate biomarkers could significantly reduce the sample sizes required to achieve statistical significance, resulting in cost benefits for future research on atherosclerotic disease. Indeed, these types of trials, which rely on an imaging based assessment of disease response, are ongoing [52]. MRI of atherosclerosis may provide useful information on how atherosclerosis responds to cholesterol lowering treatment and impacts the stability the lesions [53]. In the near future, these imaging techniques may be used to monitor disease progression in patients who are at risk of heart attack or stroke. Carotid MR can be used by vascular surgeons to improve surgical planning because the distribution and extent of disease involvement in both the CCA and ICA is better delineated by MR than by angiography. More importantly, these techniques are able to provide morphologic information that can accurately clarify the stability and configurations of a lesion regardless of the degree of stenosis. The morphology and composition of the plaque, as identified by MRI, may assist in selecting the optimal treatment – carotid endarterectomy versus angioplasty/stenting. For example, lesions with thin fibrous caps, mural thrombus formation, or highly stenotic and irregular luminal surface morphology may be better treated with surgery than stenting.


71

There are also some promising results that are based on other imaging modalities, such as ultrasonic carotid intima and media thickness measurements and their association to the incidences of myovardial infarction or stroke [54]. Future research on the association of carotid atherosclerosis morphology to myocardial infarction can result in new and important findings. With the development of new imaging and image processing techniques and the advent of new MR scanner hardware and software, current scan protocols will be converted into an efficient and economically feasible screening tool. 8. Summary Over the past decade, there have been significant advances in the field of vascular biology. Plaque burden, morphology, and composition are now considered appropriate for noninvasive quantitation and clinicians no longer rely solely on the degree of luminal narrowing to assess the severity of atherosclerotic disease. As discussed earlier in this article, MR imaging holds much potential for developing into a modality that can be used to quantitatively characterize plaque morphology in vivo. Preliminary studies have demonstrated the ability of multi-contrast MR techniques to prospectively identify the major components of human carotid plaques and to characterize the morphologic features associated with the vulnerable lesion (unstable fibrous caps, necrotic cores, intimal calcification, and intraplaque hemorrhages). Although the sample sizes of these studies are limited, continued patient recruitment and the growing number of similar or related studies performed at different institutions will hopefully establish the effectiveness of these techniques. The research results and associated technical developments described in this chapter open an exciting new era in vascular imaging and brings medical science one step closer to (1) noninvasive and prospective identification of vulnerable plaque, which would allow for more timely clinical intervention; (2) being able to quantitatively monitor the changes in disease burden and biologic markers of instability, which would enhance our understanding of the pathogenesis of the disease and the efficacy of new therapies. The high spatial resolution images presented in this article can be generated with 1.5T clinical scanners and phased-array surface coils. Currently, pulse sequences consisting of both black and bright blood techniques are recommended to characterize carotid plaque morphology in vivo. Ongoing ex vivo experiments, however, may demonstrate the utility of additional sequences like the 3D-FISP or magnetization transfer. Most importantly, the techniques described in this paper can readily be transferred, with minimal alterations, from major manufacturers to whole body scanners. Acknowledgements The authors would like to thank Andrew An Ho for his copyediting work on this manuscript. References [1] Elkins, J.S. and S.C. Johnston. Thirty-year projections for deaths from ischemic stroke in the United States. Stroke 2003; 34(9): 2109–12.

72


[2] Toole, J.F. (1990). Cerebrovascular Disorders. New York, Raven Press: 136. [3] ECST. MRC European Carotid Surgery Trial: interim results for symptomatic patients with severe (70–99). [4] NASCET. Beneficial effect of carotid endarterectomy in symptomatic patients with high-grade carotid stenosis. N Engl J Med 1991; 325(7): 445–453. [5] ACAS. Endarterectomy for asymptomatic carotid artery stenosis. JAMA 1995; 273(18): 1421–1428. [6] Moore, W.S., H.J. Barnett, et al. Guidelines for carotid endarterectomy. A multidisciplinary consensus statement from the ad hoc Committee, American Heart Association. Stroke 1995; 26(1): 188–201. [7] Eliasziw, M., J.Y. Streifler, et al. Significance of plaque ulceration in symptomatic patients with highgrade carotid stenosis. Stroke 1994; 25(2): 304–308. [8] Lusby, R.J., L.D. Ferrell, et al. Carotid plaque hemorrhage. Its role in production of cerebral ischemia. Arch Surg 1982; 117(11): 1479–1488. [9] Carr, S., A. Farb, et al. Atherosclerotic plaque rupture in symptomatic carotid artery stenosis. J Vasc Surg 1996; 23: 755–766. [10] Bassiouny, H.S., Y. Sakaguchi, et al. Juxtalumenal location of plaque necrosis and neoformation in symptomatic carotid stenosis. J Vasc Surg 1997; 26(4): 585–594. [11] Carr, S.C., A. Farb, et al. Activated inflammatory cells are associated with plaque rupture in carotid artery stenosis. Surgery 1997; 122(4): 757–63; discussion 763–4. [12] Chu, B.C., T.S. Hatsukami, et al. Determination of Carotid Artery Atherosclerotic Lesion Type and Distribution in Hypercholesterolemic Patients with Moderate Carotid Stenosis Using Non-invasive Magnetic Resonance Imaging. Stroke (in press). [13] Hatsukami TS, Ferguson MS, Beach KW, et al. Carotid plaque morphology and clinical events. Stroke 1997; 28(1):95–100. [14] Erlichman M. Surface/specialty coil devices and gating techniques in magnetic resonance imaging. Health Technol Assess Rep 1990; 3:1–23. [15] Hayes CE, Mathis CM, Yuan C. Surface coil phased arrays for high-resolution imaging of the carotid arteries. J Magn Reson Imaging 1996; 6(1):109–112. [16] Toussaint JF, Southern JF, Fuster V, Kantor HL. T2-weighted contrast for NMR characterization of human atherosclerosis. Arterioscler Thromb Vasc Biol 1995; 15(10):1533–1542. [17] Yuan C, Murakami JW, Hayes CE, et al. Phased-array magnetic resonance imaging of the carotid artery bifurcation: preliminary results in healthy volunteers and a patient with atherosclerotic disease. J Magn Reson Imaging 1995; 5(5):561–565. [18] Shinnar M, Fallon JT, Wehrli S, et al. The diagnostic accuracy of ex vivo MRI for human atherosclerotic plaque characterization. Arterioscler Thromb Vasc Biol 1999; 19(11):2756–2761. [19] Ingersleben GV, Schmiedl UP, Hatsukami TS, et al. Characterization of atherosclerotic plaques at the carotid bifurcation: correlation of high resolution MR with histology. RadioGraphics 1997; 17:1417–1423. [20] Fayad ZA, Fallon JT, Shinnar M, et al. Noninvasive in vivo high-resolution magnetic resonance imaging of atherosclerotic lesions in genetically engineered mice [see comments]. Circulation 1998; 98(15):1541–1547. [21] Mitsumori LM, Yuan C, Ferguson MS, Hatsukami TS. In vivo identification of lipid cores in advanced carotid atherosclerotic plaques by high resolution MR imaging. Circulation 2000; 102(18):II252. [22] Finn JP and Edelman RR. Black-blood and segmented k-space magnetic resonance angiography. Magn Reson Imaging Clin N Am 1993; 1(2):349–357. [23] Edelman RR, Chien D, Kim D, et al. Fast selective black blood MR imaging. Radiology 1991; 181(3):655–660. [24] Milner JS, Moore JA, Rutt BK, Steinman DA, Hemodynamics of human carotid artery bifurcations: computational studies with models reconstructed from magnetic resonance imaging of normal subjects. J Vasc Surg 1998; 28(1):143–156. [25] Steinman, DA. and B.K. Rutt. On the nature and reduction of plaque-mimicking flow artifacts in black blood MRI of the carotid bifurcation. Magn Reson Med 1998; 39(4): 635–641. [26] Hatsukami T.S., Ross R., Polissar N.L., Yuan C. Visualization of fibrous cap thickness and rupture in human atherosclerotic carotid plaque In vivo with high-resolution magnetic resonance imaging. Circulation 2000; 102(9):959–964. [27] Aoki S, Aoki K, Ohsawa S, Nakajima H, Kumagai H, Araki T. Dynamic MR imaging of the carotid wall. J Magn Reson Imaging 1999; 9(3): 420–427.


73

[28] Coombs BD, Ursell PC, Reilly LM, Rarp JH, Saloner D. Carotid Bifurcation plaque structure: high resolution MRI with histologic correlation. Abs., ISMRM Sixth Annual Meeting, Sydney, 1998. [29] Du YP, Parker DL, Davis WL, Cao G. Reduction of partial-volume artifacts with zero-filled interpolation in three-dimensional MR angiography. J Magn Reson Imaging 1994; 4(5): 733–741. [30] Merickel MB, Carman CS, Brookeman JR, Mugler JP, Brown MF, Ayers CR. Identification and 3-D quantification of atherosclerosis using magnetic resonance imaging. Comput Biol Med 1988; 18(2): 89–102. [31] Yuan C, Petty C, O’Brien KD, Hatsukami TS, Eary JF, Brown BG. In vitro and in situ magnetic resonance imaging signal features of atherosclerotic plaque-associated lipids. Arterioscler Thromb Vasc Biol 1997; 17(8):1496–1503. [32] Yuan C, Beach KW, Smith LH Jr, Hatsukami TS. Measurement of atherosclerotic carotid plaque size in vivo using high resolution magnetic resonance imaging. Circulation 1998; 98(24):2666–2671. [33] Luo Y, Polissar N, Hatsukami TS, Yuan C. Accuracy and uniqueness of three in-vivo measurements of atherosclerotic carotid plaque morphology with black blood MRI. MRM, 50:75–82, 2003. [34] Yuan C, Schmiedl UP, Maravilla KR, et al. MR imaging of human carotid plaques: assessment of reproducibility and reader variability. Proceedings of the 82th RSNA 1996; 134. [35] Kang XJ, Polissar NL, Han C, Lin E, Yuan C. Analysis of the measurement precision of arterial lumen and wall areas using high resolution magnetic resonance imaging. MRM 2000; 44:968–972. [36] Mohiaddin RH and Longmore DB. MRI studies of atherosclerotic vascular disease: structural evaluation and physiological measurements. Br Med Bull 1989; 45(4):968–990. [37] Kaufman L, Crooks LE, Sheldon PE, Rowan W, Miller T. Evaluation of NMR imaging for detection and quantification of obstructions in vessels. Invest Radiol 1982; 17(6):554–560. [38] Herfkens RJ, Higgins CB, Hricak, H, et al. Nuclear magnetic resonance imaging of atherosclerotic disease. Radiology 1983; 148(1):161–166. [39] Vinitski S, Consigny PM, Shapiro MJ, Janes N, Smullens SN, Rifkin MD. Magnetic resonance chemical shift imaging and spectroscopy of atherosclerotic plaque. Invest Radiol 1991; 26(8):703–714. [40] Small, DM and Shipley GG. Physical-chemical basis of lipid deposition in atherosclerosis. Science 1974; 185(147):222–229. [41] Gold GE, Pauly JM, Glover GH, Moretto JC, Macovski A, Herfkens RJ. Characterization of atherosclerosis with a 1.5-T imaging system. J Magn Reson Imaging 1993; 3(2):399–407. [42] Pachot-Clouard M, Vaufrey F, Darrasse L, Toussaint JF. Magnetization transfer characteristics in atherosclerotic plaque components assessed by adapted binomial preparation pulses MAGMA 1998; 7(1):9–15. [43] Kerwin W, Ferguson MS, Small R, Hatsukami TS, Yuan C. Magnetization transfer as a contrast enhancing mechanism in magnetic resonance images of human atherosclerotic plaques. Atherosclerosis 2000; 151(1):25. [44] Toussaint JF, Southern JF, Fuster V, Kantor HL. Water diffusion properties of human atherosclerosis and thrombosis measured by pulse field gradient nuclear magnetic resonance. Arterioscler Thromb Vasc Biol 1997; 17(3):542–546. [45] Sharon EC, Robert RH, J. Ross M, et al. Quantitative assessment of carotid plaque composition using multicontrast MRI and registered histology. MRM 2003; 50(6): 1199–1208. [46] Parker, DL, Yuan C, Blatter DD. MR angiography by multiple thin slab 3D acquisition. Magn Reson Med 1991; 17(2):434–451. [47] Bassiouny HS, Sakaguchi Y, Mikucki SA, et al. Juxtalumenal location of plaque necrosis and neoformation in symptomatic carotid stenosis. J Vasc Surg 1997; 26(4):585–594. [48] Gomori JM, Grossman RI, Hackney DB, et al. Variable appearances of subacute intracranial hematomas on high-field spin-echo MR. AJR Am J Roentgenol. 1988; 150; 171–178. [49] Stary HC, Chandler AB, Dinsmore RE, et al. A definition of advanced types of atherosclerotic lesions and a histological classification of atherosclerosis. A report from the Committee on Vascular Lesions of the Council on Arteriosclerosis. [50] Cai J, Hatsukami T, Ferguson M, et al. Classification of human carotid atherosclerotic lesions using in vivo multi-contrast MR imaging. Circulation 2002; 106: 1368–1373. [51] Yuan C, Zhang SX, Polissar NL, Echelard DE, Ortiz G, Davis, JW, Ellington E, Ferguson MS, Hatsukami TS. Identification of fibrous cap rupture with magnetic resonance imaging is highly associated with recent TIA or stroke. Circulation, 105:181–185; 2002. [52] Corti R, Badimon JJ, Worthley SG, Helft G, Fayad ZA, Fuster V. Statin therapy reduces burden in hu-

74


man atherosclerotic lesions: longitudinal study using noninvasive MRI. Abs. Circulation 2000; 102(18): II–459. [53] Libby P, Aikawa M, Schonbeck U. Cholesterol and atherosclerosis (review). Biochemica et Biophysica Acta 2000; 1529:299–309. [54] O’Leary DH, Polak JF, Kronmal RA, Manolio TA, Burke GL, Wolfson SK Jr. Carotid-artery intima and media thickness as a risk factor for myocardial infarction and stroke in older adults. N Engl J Med 1999; 340:14–22.


75

Intravascular Ultrasound Elastography: A Clinician’s Tool for Assessing Vulnerability and Material Composition of Plaques Radj A. BALDEWSING a , Johannes A. SCHAAR a,b , Chris L. DE KORTE a,b,c , Frits MASTIK a , Patrick W. SERRUYS a and Antonius F.W. VAN DER STEEN a,b a Thoraxcenter, Erasmus MC, Rotterdam, The Netherlands b Interuniversity Cardiology Ins. of the Netherlands (ICIN), Utrecht, The Netherlands c Clinical Physics Laboratory, Univ. Children’s Hospital, Nijmegen, The Netherlands Abstract. The material composition and morphology of the atherosclerotic plaque components are considered to be more important determinants of acute coronary ischemic syndromes than the degree of stenosis. When a vulnerable plaque ruptures it causes an acute thrombotic reaction. Rupture prone plaques contain a large lipid pool covered by a thin fibrous cap. The stress in these caps increases with decreasing thickness. Additionally, the cap may be weakened by macrophage infiltration. IntraVascular UltraSound (IVUS) elastography might be an ideal technique to assess the presence of lipid pools and to identify high stress regions. Elastography is a technique that assesses the local elasticity (strain and modulus) of tissue. It is based on the principle that the deformation of tissue by a mechanical excitation is a function of its material properties. The deformation of the tissue is determined using ultrasound. For intravascular purposes, the intraluminal pressure is used as the excitation force. The radial strain in the tissue is obtained by cross-correlation techniques on the radio frequency signals. The strain is color-coded and plotted as a complimentary image to the IVUS echogram. IVUS elastography, and IVUS palpography (which uses the same principle but is faster and more robust), have been extensively validated using simulations and by performing experiments in vitro and in vivo with diseased arteries from animals and humans. Strain was shown to be significantly different in various plaque types (absent, fatty, fibrous or calcified). A high strain region with adjacent low strain at the lumen vessel-wall boundary has 88% sensitivity and 89% specificity for detecting vulnerable plaques. High strain regions at the lumen plaque-surface have 92% sensitivity and 92% specificity for identifying macrophages. Furthermore, the incidence of vulnerable-plaque-specific strain patterns in humans has been related to clinical presentation (stable angina, unstable angina or acute myocardial infarction) and the level of C-reactive protein. In conclusion, the results obtained with IVUS (strain and modulus) elastography/palpography, show the potential of the technique to become a unique tool for clinicians to assess the vulnerability and material composition of plaques. Keywords. Tissue characterisation, intravascular, ultrasound, elastography, vulnerable, plaque, stress, strain, modulus

76

R.A. Baldewsing et al. / Intravascular Ultrasound Elastography

1. Introduction IntraVascular UltraSound (IVUS) is the only commercially available clinical technique providing real-time cross-sectional images of the coronary artery in patients [1]. IVUS provides information on the severity of the stenosis and the remaining free luminal area. Furthermore, calcified and non-calcified plaque components can be identified. Although many investigators studied the value of IVUS to identify the plaque composition, identification of fibrous and fatty plaque components remains limited [2,3]. IVUS RadioFrequency (RF)-based tissue identification strategies appear to have better performance [3,4]. However, none of them is yet capable of providing sufficient spatial and parametric resolution to identify a lipid pool covered by a thin fibrous cap. Identification of different plaque components is of crucial importance to detect the high-risk, vulnerable plaque since these are characterised by an eccentric plaque with a large lipid pool shielded from the lumen by a thin fibrous cap [5,6]. Inflammation of the cap by macrophages further increases the vulnerability of these plaques [7]. The mechanical properties of fibrous and fatty plaque components are different [8–10]. Furthermore, fibrous caps with inflammation by macrophages are weaker than caps without inflammation [11]. The stress that is applied on an artery by the pulsating blood pressure must balance the circumferentially directed load integrated over the whole arterial wall. To maintain the connection between mechanically different tissue structures (like soft lipid pools and stiff fibrous caps) during arterial deformation, relatively soft regions will therefore carry only a fraction of the total circumferential load and the surrounding stiffer material a greater portion [12,13]. This mechanism causes circumferential stress concentrations in and around the stiff cap, which will rupture if the cap is unable to withstand this stress. This increased circumferential stress will result in an increased radial deformation (strain) of the tissue due to the incompressibility of the material. Therefore, methods that are capable of measuring the radial strain provide information about plaques that may influence clinical decision-making. In 1991, Ophir et al. [14] proposed a method to measure the elasticity (strain and modulus) of biological tissues using ultrasound. The tissue was deformed by externally applying a stress on it. Different strain values were found in tissues with different material properties. Implementing this method for intravascular purposes has potential to identify the vulnerable plaque by (i) identification of different plaque components and (ii) detection of high strain regions. This chapter discusses, the technique behind, the method for and the validation of IVUS elastography, which is differentiated into IVUS strain elastography/palpography when the strain is imaged and IVUS modulus elastography when the modulus is imaged.

2. Ultrasound Elastography 2.1. The Movement Begins In 1991, Ophir and colleagues [14,15] developed an imaging technique called elastography, which is based on (quasi-) static deformation of a linear elastic, isotropic material. The tissue under inspection is deformed by applying stress (i.e., force normalized by area) on a part of its boundary. The resulting distribution of strain (i.e., length of a small block of tissue after deformation, divided by its length before deformation) depends upon


77

(i) the distribution of the tissue’s material properties (Young’s modulus and Poisson’s ratio) and (ii) the displacement or stress conditions on the remaining tissue boundaries. The Young’s modulus E (kPa) is a material property, which can be interpreted as the ratio between the stress S (kPa) (tensile or compressive) enforced upon a small block of tissue and its resulting strain (elongation or compression). The Poisson’s ratio is also a material property and it quantifies a material’s local volumetric compressibility. The resulting strain is determined, directly or indirectly using displacement, with ultrasound using two pairs of ultrasound signals, one signal obtained before and the other after deformation [16]. The method was developed for detection and characterization of tumors in breast. Nowadays, this principle is applied on many other biological objects [17] including prostate, kidney, liver, myocardium and skin. Recently, some preliminary (simulation) results have been reported on the application of this principle to non-invasively assess elastic properties of superficial peripheral arteries like carotid, brachial, radial and femoral [18–21]. Although Ophir et al. never explored the quasi-static approach for intravascular purposes this approach seems to be the most fruitful concept. In this application, beside knowledge of the material properties of the different plaque components, the strain in itself may be an excellent diagnostic parameter. Furthermore, in intravascular applications, the arterial deformation is naturally present and is caused by the pulsatile blood pressure. Eventually, user-controlled deformation is possible by using a compliant intravascular balloon [22]. 2.2. IVUS Strain Elastography/Palpography The principle of IVUS strain elastography is illustrated in Fig. 1. An ultrasound image of a vessel-phantom with a stiff vessel wall and a soft eccentric plaque is acquired at a low pressure. In this case, there is no difference in echogenicity between the vessel wall and the plaque resulting in a homogeneous IVUS echogram. A second acquisition at a higher intraluminal pressure (pressure differential is approximately 1 mm Hg) is obtained. The radial strain elastogram is plotted as a complimentary image to the IVUS echogram. The elastogram reveals the presence of an eccentric region with increased strain values thus identifying the soft eccentric plaque. The differences in strategies to perform IVUS strain elastography (i.e. assess the local deformation of the tissue) are due to (i) the way of detecting the strain and (ii) the type of source that deforms the vascular tissue. The principle of IVUS strain palpography is similar to the principle of IVUS strain elastography. There are two minor differences that make palpography faster and more robust and, therefore, more suited for real-time in vivo applications. Firstly, palpography restricts its region of interest to the innermost layer of the arterial wall (first 500 micrometer), making it faster. Secondly, it uses a slightly larger amount of ultrasound signal making it more robust but at the expense of spatial resolution [23–25]. 2.3. Implementation of the Technique Typically in IVUS strain elastography/palpography, intraluminal pressure differences in the order of 1–5 mm Hg are used. The strain, induced by this pressure differential in vascular tissue is in the order of 2%. This means that a small block of tissue with an initial length of 100 micrometer will be deformed to 98 micrometer. To differentiate between strain levels, sub-micron estimation of the tissue displacement is required.

78


Figure 1. Principle of intravascular ultrasound (IVUS) strain elastography measurement procedure. An intravascular ultrasound catheter is inserted into an object, in this case a vessel-mimicking phantom with a soft plaque (A). Next, at two different intraluminal pressures (B) an IVUS echogram is acquired (C and D). In each echogram, the grey circle indicates the catheter-tip of 1.1 mm diameter. Finally, the radial strain in the tissue is determined using cross-correlation processing on the acquired radio-frequency data. This information is plotted (strain elastogram) as an additional image to the IVUS echogram (E). In this example, the eccentric soft plaque of a vessel-mimicking phantom is clearly visible between 4 and 8 o’clock in the strain elastogram whereas this plaque cannot be identified from the IVUS echograms.

2.3.1. Envelope Based Two groups worked on intravascular elastography using the envelope of the ultrasound signal. Talhami et al. [26] introduced a technique to assess the strain that is based on the Fourier scaling property of the signals and uses the chirp Z-transform (i.e. a mathematical transformation of the ultrasound signal into another signal to make it suitable for some desired mathematical manipulations). The scaling property is a direct estimator of the strain in the tissue. The chirp-Z transform was determined from the envelope signal to overcome decorrelation of the Radio-Frequency (RF)-signal due to deformation of the tissue. The result was displayed as a color-coded ring superimposed on the IVUS echogram of the vessel. Initial results on vascular tissue in vitro and in vivo were described. Although the technique seems relatively easy to implement, it was not further developed and validated. Ryan and Foster [27] developed a technique to estimate displacement by using speckle tracking (i.e. following the movement of a signal pattern in series of images) in video signals. The strain can be determined from these displacement estimates. The technique was tested using vessel-mimicking phantoms. It was shown that in a phantom that was partly made of soft and partly made of stiff material the displacement in soft material was larger than the displacement in stiff material. An advantage of envelope-based methods is the fact that the correlation function is smoother than the RF-based correlation function. This prevents ‘peak hopping’, meaning that the correlation function is maximized around the wrong peak. This makes the method less noise sensitive. Furthermore, the video signal is commonly available from any commercial echo system. A disadvantage is the limited resolution and the low sensitivity of the method for low strain values. Since small tissue strains are expected for intravascular applications, and the arterial wall is relatively thin, it is expected that the use of the high


79

frequency RF-signal will greatly improve the resolution. Based on work of Varghese and Ophir [28], a smaller variance of the strain estimate is expected using RF data instead of envelope data. 2.3.2. Radio Frequency Based Shapo et al. [29,30] developed a technique based on cross-correlation of A-lines. The group proposes a large deformation to maximize the signal-to-noise ratio of the displacement and strain estimation. Since the deformation that is needed is larger than the deformation that occurs in arteries in vivo, the artery is deformed using a non-compliant balloon that is inflated up to 8 atmospheres. Large displacement will decorrelate the ultrasound signals to such an extent that correlation detection is unreliable. For this reason, the cross-correlation is calculated in several intermediate steps of intraluminal pressure. For detection, they use a phase-sensitive speckle tracking technique. The technique was demonstrated in simulations and tissue mimicking phantoms. This group presented data on a compliant balloon containing an intravascular catheter [31,32]. The compliant balloon is inflated to 2 atmospheres to obtain strain values up to 40%. This method was tested in phantoms and in vitro. The phantom (with one half soft and one half stiff material) revealed strains from 20–40% in the soft part and strains lower than 20% in the stiff part. These results were corroborated by finite element analysis. In vitro, low strain values (7%) were found for fibrous tissue in the human femoral artery and high strain values (35%) were found in thrombus in a rabbit aorta. De Korte et al. [33,34] incorporated ‘correlation-based’ elastography [14] for intravascular purposes. The vascular tissue is strained by different levels of the intraluminal pressure. The local displacement of the tissue is determined using cross-correlation analysis of the gated RF-signals (Fig. 2). A cross-correlation function between two signals will have its maximum if the signals are not shifted with respect to each other. If a shift between the signals is present, the peak of the cross-correlation function is found at the position representing the displacement of the tissue. For each angle, the displacement of the tissue at the lumen vesselwall boundary is determined. Next, the displacement of the tissue D micrometer from the vessel-wall boundary is determined. The strain of the tissue can be calculated by dividing the differential displacement (displacement of tissue at boundary – displacement of tissue in wall) by the distance between these two locations (D micrometer). The strain for each angle is color-coded and plotted as a ring (strain palpogram) on the IVUS echogram at the lumen vessel-wall boundary [23,24,35]. If the strain is determined for multiple regions per angle, a two-dimensional image of the strain can be constructed. This additional image to the IVUS echogram is called an IVUS strain elastogram. For palpography, the value for D is approximately 400 micrometers and for elastography approximately 200 micrometers. The method was validated using vessel phantoms with the morphology of an artery with an eccentric soft or stiff plaque [36]. The plaque could be clearly identified from the vessel-wall using the strain elastogram, independently of the echogenicity contrast between vessel-wall and plaque. The technique was validated in vitro and tested in atherosclerotic animal models and during interventions in patients (as discussed later). This cross-correlation based technique is especially suited for strain values smaller than 2.5% (for a signal segment in the order of 10 wavelenghts). These strain values are present during in vivo acquisitions when only a part of the heart cycle is used to strain the tissue [37]. The maximum strain that will be present between the systolic and diastolic pressure is much higher. In that case, another approach that takes

80


Figure 2. Principle of time delay estimation using the peak of the cross-correlation coefficient function. In the upper part, two segments (A and B) of the pre-deformation (solid line) and post-deformation (dotted, and preshifted for better visual comparison) radio-frequency (RF) signals are shown. Both segments start at a different position in the tissue. For each segment, the cross-correlation coefficient function between the two signals is computed (C and D). The functions show a decreasing position of the peak with increasing echodepth. The difference in peak position represents the differential displacement.

into account the change in shape of the signals can be applied. This ‘local scaling factor estimation’ technique [38] has been recently described for intravascular purposes [39] and has proven to be more robust to large deformations. The signal after compression is processed as a delayed and scaled replica of the signal before deformation. An adaptive strain estimation method based on the computation of local scaling factors has been applied to compute strain elastograms of cryogel vessel-mimicking phantoms and of a freshly excised human carotid artery using a 30-MHz mechanical rotating single element ultrasound scanner (ClearView, CVIS, Boston Scientific Corp.) [40,41]. Recently, more


81

Figure 3. IVUS strain elastography in vitro of a human femoral artery and corresponding histology. (A) IVUS echogram. (B) IVUS strain elastogram superimposed on the IVUS echogram. Histology: (C) Collagen, (D) Smooth muscle cells and (E) Macrophages. The elastogram reveals that the plaque contains a region of high strain between 1 and 4 o’clock. Histology shows that this region is heavily infiltrated by macrophages and lacks smooth muscle cells and collagen. Furthermore, the remaining plaque region between 4 and 1 o’clock shows low strain and, at this region, histology reveals that the plaque contains much smooth muscle cells and collagen but no macrophages.

groups have started to perform IVUS strain elastography using a rotating single element scanner [42–44].

3. IVUS Strain Elastography/Palpography: Results 3.1. In Vitro Validation on Human Arteries De Korte et al. [45] performed a validation study on excised human coronary (n = 4) and femoral (n = 9) arteries (Fig. 3) to investigate the capability of IVUS strain elastography to characterize different plaque components. Data were acquired at room temperature at intraluminal pressures of 80 and 100 mm Hg. Coronary arteries were measured using a solid state 20-MHz array catheter (EndoSonics, Rancho Cordova, CA, USA). Femoral arteries were investigated using a single element 30-MHz catheter (DuMed/ EndoSonics, Rijswijk, The Netherlands) that was connected to a modified motor unit (containing the pulser and receiver and a stepper-motor to rotate the catheter). The RF-data was stored and processed off-line. The visualized segments were stained on the presence of collagen, smooth muscle cells and macrophages. Matching of elastographic data and histology was performed using the IVUS echogram. The cross-sections were segmented in regions (n = 125) based on the strain value on the elastogram. The dominant plaque types in these regions (fibrous, fibro-fatty or fatty) were obtained from histology and correlated with the average strain and echo-intensity. Mean strain values of 0.27%, 0.45% and

82


0.60% were found for fibrous, fibro/fatty and fatty plaque components. The strain for the three plaque types as determined by histology differed significantly (p = 0.0002). This difference was independent on the type of artery (coronary or femoral) and was mainly evident between fibrous and fatty tissue (p = 0.0004). The plaque types did not reveal echo-intensity differences in the IVUS echogram (p = 0.882). Conversion of the strain into an Incremental Stress-Strain Modulus (denoted as Eissm) was performed using the formula Eissm = (dP/2)/(meanstrain); where dP is the intraluminal pressure differential and ‘meanstrain’ is the mean of the strain around the circumference of the first layer of vessel wall. Because the pressure differential and thus the stress are only known at the boundary between lumen and vessel-wall and due to non-linearity of this parameter this gives only a first order approximation of the modulus. The resulting Eissm was 493 kPa, 296 kPa and 222 kPa for fibrous, fibro/fatty and fatty plaques. Although these values are higher than values measured by Lee et al. [10], the ratio between fibrous and fatty material is similar. Since fibrous and fatty tissue resulted in different strain values and high strain values often co-localised with increased concentrations of macrophages, these results reveal the potential of identification of the vulnerable plaque. Recently, Schaar et al. [46] performed a study on human coronary arteries to quantify the predictive value of IVUS strain elastography to detect the vulnerable high-risk plaque. Postmortem coronary arteries (n = 24) were investigated with IVUS strain elastography using a 20-MHz phased array catheter (JOMED, Rancho Cordova, CA) connected to an echo apparatus (JOMED, Rancho Cordova, CA). Subsequently they were processed for histology. In histology, a vulnerable plaque was defined as a plaque consisting of a thin cap ( 0.49) elastic material. In those cases the constitutive relation contains only one material parameter, namely the Young’s modulus. Finally, the deformation model and the measured displacement/strain components are used to compute the modulus distribution by a ‘direct solution approach’ or by an ‘iterative solution approach’. • ‘Direct’: In the direct solution approach the measured displacement/strain data are plugged in the deformation equations, which are mathematically manipulated so that the moduli can be considered and expressed as the unknowns. Next, the moduli are computed using a discretization [64] or numerical integration of the manipulated deformation equations [65,66]. • ‘Iterative’: In the iterative approach, a Finite Element (computer) Model (FEM) is taken as the deformation model. The FEM fills the space of the tissue with a mesh that consists of small elements (e.g. triangles, bricks) and each element is given a constitutive relation, i.e. Young’s modulus. Next, an initial modulus value for each element defined. Finally the modulus value of each individual mesh element or groups of mesh elements in the FEM are iteratively changed such that the computed FEM deformation output eventually closely matches the measured deformation (displacement/strain data). This matching is fully automatic performed by a minimization algorithm [67]. Much research has focussed on applying these two approaches on non-vascular tissue geometries like a two dimensional cross-section of a homogeneous rectangular medium with a circular or rectangular inclusion [64,66–72], or on a breast or brain part [73,74], or heart [75]. To date, only a few groups have investigated the inverse problem for vascular geometries. Most of them used an adjusted iterative reconstruction method [76–79] and [80], some others an adjusted direct reconstruction method [81,82]. All groups encountered difficulties in computing a modulus elastogram (related to uniqueness and continuity), which may be caused by noisy measurements, a limited number of measured displacement/strain components, type of boundary data [83], using an inadequate deformation model for the tissue, non-uniqueness of the inverse problem [84] and converging to non-optimal local minima by the minimization algorithm. 4.4. IVUS Modulus Elastography Recently, Baldewsing et al. [85] focused on solving the intravascular inverse problem for atherosclerotic vascular geometries using an iterative solution approach. Thereby, they used the arterial radial strain measured with IVUS strain elastography and an a priori

88


Figure 7. Parametric finite element model for a vulnerable plaque. (A) Each circle is parameterized by its center (X,Y) and a radius R. The dynamic control points P, Q, R, S, T, and U are used to define the three plaque component regions. (B) Finite element mesh regions corresponding to geometry in A. Each finite element in a region has the same material property values as other elements in that region. Letters L and c denote respectively lipid and cap.

parametric-plaque-geometry computer model. Their motivation for using a priori information is threefold. Firstly, they want to compute a modulus image of an atherosclerotic plaque that is diagnostically useful and easy to interpret in clinical settings. Secondly, the computation should suffer at least as possible from convergence problems. Finally, they want to be able to investigate and quantify difficulties (uniqueness and continuity) and limitations for computing a modulus elastogram of plaques in a structured manner. There iterative solution approach is especially suited for ‘thin-cap fibroatheromas’ (TCFA) [52,53]. Their approach is as follows: the deformation output calculated with a Parametric Finite Element Model (PFEM) representation of a TCFA, is matched to the plaque’s radial strain as measured with IVUS strain elastography. The PFEM uses only six morphology and three material composition parameters, but is still able to model a variety of these TCFA’s. The computed Young’s modulus image of the TCFA shows both the morphology and Young’s modulus value of three main plaque components, namely lipid, cap and media and should therefore be easy interpretable. In the next three subsections, the main parts of their iterative solution approach are discussed, namely the PFEM for a TCFA, the used deformation model and used minimization algorithm. • ‘PFEM Geometry for a Plaque’: An idealized ‘thin-cap fibroatheroma’ [52] is used as a model for a plaque and is an extension of the PFEM model used by Loree et al. [12]. The PFEM geometry consists of a media area containing a lipid pool, which is covered by a fibrous cap. The borders of the lipid, cap and media areas are defined using circles (Fig. 7). Lipid is defined by region QTQ, cap by region PQTSP, and media by the remaining area. Each circle is parameterised by its center with Cartesian coordinates (X,Y) and radius R. Resulting in a total of six morphology parameters. • ‘Material Deformation Model’: Baldewsing et al. [86] used coronary arteries (n = 5) to demonstrate that radial strain elastograms measured in vitro using IVUS strain elastography could be simulated with a finite element model. Their material


89

deformation model treated the arterial tissue as a linear elastic, isotropic, planestrain, nearly incompressible material with a Poisson’s ratio of 0.4999 [55].The finite element model geometry and material properties were determined using the IVUS echogram and histology (collagen, smooth muscle cells and macrophages). The agreement between a simulated and measured elastogram was quantified by comparing features of high strain regions. Statistical test showed that there was no significant difference between simulated and corresponding measured elastograms in location, surface area and mean strain value of a high strain region. The same material deformation model is used for the PFEM. Lipid, cap and media region are assumed to have a constant Young’s modulus value EL, EC, and EM, respectively. This results in a total of only three material composition parameters. The PFEM radial strain deformation is computed using the finite element package SEPRAN (Sepra Analysis, Technical University Delft, The Netherlands); the catheter center is taken as origin. This radial strain field is called a PFEM strain elastogram. The whole process from defining the PFEM morphology and material composition parameters up to the calculation of the PFEM strain elastogram is fully automatic. • ‘Minimization Algorithm’: The modulus elastogram of a plaque is determined by a minimization algorithm. This algorithm tries to find values for the six morphology and three modulus parameters of the PFEM such that the corresponding PFEM strain elastogram ‘looks similar’ to the measured IVUS strain elastogram. Theire similarity is quantified as the Root-Mean-Squared (RMS) error between PFEM strain elastogram and measured IVUS strain elastogram. The sequentialquadratic-programming minimization algorithm fully automatic searches a local minimum of the RMS by iteratively updating the nine PFEM parameters. Each update gives a lower RMS error. The algorithm stops when the either the RMS itself or each of a few consecutive RMS values is below a threshold value. When the resulting final PFEM strain elastogram has qualitatively enough strain pattern features in common with the measured IVUS strain elastogram, the corresponding modulus elastogram is considered to be a good approximation of the material composition of the investigated plaque.

5. IVUS Modulus Elastography: Preliminary Results Baldewsing et al. [87,88] have shown the feasibility of their approach by successfully applying their IVUS modulus elastography method on radial strain elastograms of vulnerable plaques that were (a) simulated, (b) measured in vitro and (c) measured in vivo in a patient. Two computer-simulated plaques, a plaque-mimicking phantom and two human coronary plaques (in vitro and in vivo) were used. Finite element models were used to simulate strain elastograms for two plaque geometries, both having a lipid pool covered by a cap; one geometry was defined by circles, the other by tracing arterial histology. For their in vitro phantom and coronary artery, strain elastograms were processed from radiofrequency data obtained with a 20-MHz 64-element phased array IVUS catheter. For the patient’s case, multiple in vivo strain elastograms, obtained during the diastolic phase of a cardiac cycle were catheter motion was minimal, were averaged into one in vivo compounded strain elastogram to increase the signal-to-noise ratio. All their computed mod-

90


Figure 8. Computing an IVUS modulus elastogram from an IVUS strain elastogram that was measured in vivo from a patient with a vulnerable plaque. (A) IVUS echogram. (B) In vivo measured compounded strain elastogram. (C) Computed Young’s modulus elastogram in kPa. (D) PFEM strain elastogram computed from C.

ulus elastograms approximated the geometry and material properties of the real plaque composition. Figure 8 shows computation of an IVUS modulus elastogram from the in vivo measured compounded IVUS strain elastogram of the patient. The echogram (Fig. 8A) reveals the presence of a large eccentric plaque between 10 and 5 o’clock, but cannot discriminate between the possible cap and lipid component of the plaque. The in vivo measured compounded IVUS strain elastogram (Fig. 8B) suggests the presence of a soft lipid pool covered by a stiff cap by means of a typical high radial strain region at the shoulders of the plaque and mechanical shadowing, which causes the low strain at the center of the plaque. The pressure differential was 1 mm Hg. The computed Young’s modulus elastogram (Fig. 8C) with (EL = 1, EC = 118, and EM = 111 kPa) is a likely candidate for the unknown material composition of the plaque, since the measured compounded IVUS strain elastogram (Fig. 8B) and the PFEM strain elastogram (Fig. 8D) show two co-localizing high strain regions and mechanical shadowing.

6. Discussion Identification of plaque components and the proneness of a lesion to rupture is a major issue in interventional cardiology. Intravascular ultrasound (IVUS) echography is a real-time, clinical available technique capable of providing cross-sectional images and identifying calcified plaque components. Since IVUS strain elastography only requires ultrasound data sets that are acquired at different levels of intraluminal pressure, it can


91

be realised using conventional catheters. It has been shown by several research groups that IVUS strain elastograms of vessel-like phantoms and arteries in vitro and in vivo can be produced. Furthermore, the feasibility of IVUS strain elastography in vivo in animals and patients was demonstrated. Currently, there is no clinical available technique capable of identifying the rupture prone plaque. Identification of these plaques is of paramount importance to investigate the underlying principle of plaque rupture, the effectiveness of pharmaceutical treatments and, on the long-term, preventing sudden cardiac deaths. IVUS strain elastography has proven to be able to identify the rupture prone plaque in vitro with high sensitivity and specificity and in vivo experiments demonstrate the power to identify fibrous and fatty plaque components. Therefore, IVUS strain elastography is one of the first techniques that can be applied in patients to assess the vulnerability of plaques. Since IVUS strain palpography is a faster and more robust technique, introduction of this technique in the catheterisation laboratory may be easier. Although IVUS strain palpography reveals no information on the composition of material deeper in the plaque, it may be a powerful technique to identify the weak spots in an artery. If a plaque will rupture, this rupture will start at the lumen vessel-wall boundary and this region is imaged with IVUS strain palpography. Recently 3D IVUS strain palpography was used in vivo to detect vulnerable-plaque-specific strain patterns in human coronary arteries and this study showed that the total number of these patterns correlated both with the clinical presentation and the level of C-reactive protein. Further studies should be performed to assess the potential role of this technique to identify patients at risk of future clinical events. Quantitative monitoring of atherosclerosis to study the effect of pharmaceutical therapies aimed at stabilizing plaques, e.g. by stiffening or reducing the lipids [89], requires a technique that provides quantitative information about the material properties of plaque components. IVUS modulus elastography seems most suited for this application, since it images the modulus of tissue, which is a material property and therefore its value is independent of the morphology and location of plaque constituents in the arterial wall, in contrast to strain. Since a large modulus contrast between soft and stiff plaque components exists, differentiation between them is straightforward. The feasibility of IVUS modulus elastography using an iterative solution approach and an a priori plaque geometry model was show using simulations and in vitro using a vessel phantom and a human coronary artery that both contained a TCFA. However, plaques can also have a complex, heterogeneous material composition consisting of mixture of plaque components, like lipids, fibrotic tissue, calcified nodules or tissues weakened by inflammation due to macrophage infiltration. Imaging of such complexes requires a different approach, which doesn’t enforce restrictions on the possible plaque constituents and their morphologies. A direct solution approach is an option. However, the following hybrid approach seems also fruitful: First an iterative approach with a priori plaque information is applied, like Baldewsing et al. [85] did, to obtain an global approximation of the modulus distribution. Next, this distribution is used as initilization in the iterative approach reported by Soulami et al. [79] that computes for each small individual finite element in the plaque a modulus. Although their approach does not require a priori information, the large number of moduli to be computed might compromise a successful convergence of the minimization algorithm. Finally, a hybrid construction between a direct and iterative approach should not be ruled out. To realize IVUS modulus elastography as a practical tool for clinical application, the following research still has to be performed (i) developing a solution approach for an arbitrary plaque that is robust

92


(i.e. the influence of catheter position and measurement noise should not significantly influence the outcome of the modulus elastogram), (ii) validation in vitro and/or in vivo and finally, (iii) investigating the need for and possibility of real-time application. All the results obtained up to now show the potential of IVUS (strain and modulus) elastography/palpography to become a unique tool for clinicians to detect vulnerable high-risk plaques, selecting proper interventional procedures and monitoring atherosclerosis.

7. Conclusion Intravascular ultrasound (IVUS) strain elastography has proven to be a technique capable of providing information about the vulnerability and material composition of plaques. In vivo studies in animals and patients demonstrate that IVUS strain palpography may develop into a clinical available tool to identify the rupture prone plaque. Future studies will reveal the potential of IVUS modulus elastography to quantitatively monitor atherosclerosis.

Acknowledgement The research discussed in this chapter has been funded by grants from the Dutch Technology Foundation (STW), the Netherlands Organization for Scientific Research (NWO), Deutsche Herzstiftung (DHS), Dutch Heart Foundation (NHS) and JOMED.

References [1] MacNeill, B.D., Lowe, H.C., Takano, M., Fuster, V., and Jang, I.K., Intravascular modalities for detection of vulnerable plaque: current status, Arterioscler Thromb Vasc Biol, Vol. 23, No. 8, pp. 1333–42, 2003. [2] Prati, F., Arbustini, E., Labellarte, A., Bello, B.D., Sommariva, L., Mallus, M.T., Pagano, A., and Boccanelli, A., Correlation between high frequency intravascular ultrasound and histomorphology in human coronary arteries, Heart, Vol. 85, pp. 567–570, 2001. [3] Komiyama, N., Berry, G., Kolz, M., Oshima, A., Metz, J., Preuss, P., Brisken, A., Moore, M., Yock, P., and Fitzgerald, P., Tissue characterization of atherosclerotic plaques by intravascular ultrasound radiofrequency signal analysis: An in vitro study of human coronary arteries, American Heart Journal, Vol. 140, pp. 565–574, 2000. [4] Hiro, T., Fujii, T., Yasumoto, K., Murata, T., Murashige, A., and Matsuzaki, M., Detection of fibrous cap in atherosclerotic plaque by intravascular ultrasound by use of color mapping of angle-dependent echo-intensity variation, Circulation, Vol. 103, No. 9, pp. 1206–11, 2001. [5] Falk, E., Shah, P.K., and Fuster, V., Coronary plaque disruption, Circulation, Vol. 92, No. 3, pp. 657–671, 1995. [6] Davies, M.J., The pathophysiology of acute coronary syndromes, Heart, Vol. 83, No. 3, pp. 361–366, 2000. [7] Moreno, P.R., Falk, E., Palacios, I.F., Newell, J.B., Fuster, V., and Fallon, J.T., Macrophage infiltration in acute coronary syndromes. Implications for plaque rupture, Circulation, Vol. 90, No. 2, pp. 775–8, 1994. [8] Loree, H.M., Tobias, B.J., Gibson, L.J., Kamm, R.D., Small, D.M., and Lee, R.T., Mechanical properties of model atherosclerotic lesion lipid pools, Arterioscler Thromb, Vol. 14, No. 2, pp. 230–4, 1994. [9] Loree, H.M., Grodzinsky, A.J., Park, S.Y., Gibson, L.J., and Lee, R.T., Static circumferential tangential modulus of human atherosclerotic tissue, J Biomech, Vol. 27, No. 2, pp. 195–204, 1994. [10] Lee, R.T., Richardson, G., Loree, H.M., Gordzinsky, A.J., Gharib, S.A., Schoen, F.J., and Pandian, N., Prediction of mechanical properties of human atherosclerotic tissue by high-frequency intravascular ultrasound imaging, Arterioscler Thromb, Vol. 12, pp. 1–5, 1992.


93

[11] Lendon, C.L., Davies, M.J., Born, G.V.R., and Richardson, P.D., Atherosclerotic plaque caps are locally weakened when macrophage density is increased., Atherosclerosis, Vol. 87, pp. 87–90, 1991. [12] Loree, H.M., Kamm, R.D., Stringfellow, R.G., and Lee, R.T., Effects of fibrous cap thickness on peak circumferential stress in model atherosclerotic vessels, Circ Res, Vol. 71, No. 4, pp. 850–858, 1992. [13] Richardson, P.D., Davies, M.J., and Born, G.V.R., Influence of plaque configuration and stress distribution on fissuring of coronary atherosclerotic plaques, The Lancet, Vol. 2, pp. 941–944, 1989. [14] Ophir, J., Cespedes, I., Ponnekanti, H., Yazdi, Y., and Li, X., Elastography: a quantitative method for imaging the elasticity of biological tissues, Ultras Imag, Vol. 13, No. 2, pp. 111–34, 1991. [15] Céspedes, E.I., Ophir, J., Ponnekanti, H., and Maklad, N., Elastography: elasticity imaging using ultrasound with application to muscle and breast in vivo, Ultras Imag, Vol. 17, pp. 73–88, 1993. [16] Céspedes, E.I., Huang, Y., Ophir, J., and Spratt, S., Methods for estimation of subsample time delays of digitized echo signals, Ultras Imag, Vol. 17, pp. 142–171, 1995. [17] Konofagou, E.E., Quo vadis elasticity imaging?, Ultrasonics, Vol. 42, No. 1–9, pp. 331–6, 2004. [18] Mai, J.J., and Insana, M.F., Strain imaging of internal deformation, Ultrasound Med Biol, Vol. 28, No. 11–12, pp. 1475–84, 2002. [19] Lupotti, F.A., Mai, J.J., Pellot-Barakat, C., and Insana, M.F., Vascular elasticity from regional displacement estimates, Proceedings of the 2003 IEEE International Ultrasonics Symposium, pp. 1895–1898, 2003. [20] Weitzel, W.F., Kim, K., Rubin, J.M., Xie, H., Chen, X., and O’Donnell, M., Peripheral arterial strain imaging using external pressure equilization, Proceedings of the Second International Conference on the Ultrasonic Measurement and Imaging of Tissue Elasticity, p. 74, 2003. [21] Maurice, R.L., Ohayon, J., Fretigny, Y., Bertrand, M., Soulez, G., and Cloutier, G., Noninvasive vascular elastography: theoretical framework, IEEE Trans Med Imaging, Vol. 23, No. 2, pp. 164–180, 2004. [22] Sarvazyan, A.P., Emelianov, S.Y., and Skovorada, A.R. Intracavity device for elasticity imaging. US patent; 1993. [23] Doyley, M., Mastik, F., de Korte, C.L., Carlier, S., Cespedes, E., Serruys, P., Bom, N., and van der Steen, A.F.W., Advancing intravascular ultrasonic palpation towards clinical applications, Ultrasound in Medicine and Biology, Vol. 27, No. 11, pp. 1471–1480, 2001. [24] Céspedes, E.I., de Korte, C.L., and van der Steen, A.F.W., Intraluminal ultrasonic palpation: assessment of local and cross-sectional tissue stiffness, Ultrasound in Medicine and Biology, Vol. 26, No. 3, pp. 385–396, 2000. [25] Schaar, J.A., de Korte, C.L., Mastik, F., Baldewsing, R., Regar, E., de Feyter, P., Slager, C.J., van der Steen, A.F.W., and Serruys, P.W., Intravascular palpography for high-risk vulnerable plaque assessment, Herz, Vol. 28, No. 6, pp. 488–95, 2003. [26] Talhami, H.E., Wilson, L.S., and Neale, M.L., Spectral tissue strain: a new technique for imaging tissue strain using intravascular ultrasound, Ultrasound in Medicine and Biology, Vol. 20, No. 8, pp. 759–772, 1994. [27] Ryan, L.K., and Foster, F.S., Ultrasonic measurement of differential displacement and strain in a vascular model, Ultras Imag, Vol. 19, pp. 19–38, 1997. [28] Varghese, T., and Ophir, J., Characterization of elastographic noise using the envelope of echo signals, Ultrasound in Medicine and Biology, Vol. 24, No. 4, pp. 543–55, 1998. [29] Shapo, B.M., Crowe, J.R., Erkamp, R., Emelianov, S.Y., Eberle, M., and O’Donnell, M., Strain imaging of coronary arteries with intraluminal ultrasound: experiments on an inhomogeneous phantom, Ultras Imag, Vol. 18, pp. 173–191, 1996. [30] Shapo, B.M., Crowe, J.R., Skovoroda, A.R., Eberle, M., Cohn, N.A., and O’Donnell, M., Displacement and strain imaging of coronary arteries with intraluminal ultrasound, IEEE trans UFFC, Vol. 43, No. 2, pp. 234–246, 1996. [31] Choi, C.D., Skovoroda, A., Emelianov, S., and O’Donnell, M. Strain imaging of vascular pathologies using a compliant balloon catheter. IEEE Ultrasonics Symposium, Puerto Rico, USA: 1771–1774; 2000. [32] Choi, C.D., Skovoroda, A.R., Emelianov, S.Y., and O’Donnell, M., An integrated compliant balloon ultrasound catheter for intravascular strain imaging, IEEE Trans Ultrason Ferroelectr Freq Control, Vol. 49, No. 11, pp. 1552–60, 2002. [33] de Korte, C.L., Cespedes, E.I., van der Steen, A.F., Pasterkamp, G., and Bom, N., Intravascular ultrasound elastography: assessment and imaging of elastic properties of diseased arteries and vulnerable plaque, Eur J Ultrasound, Vol. 7, No. 3, pp. 219–24, 1998. [34] de Korte, C.L., van der Steen, A.F.W., Cepedes, E.I., Pasterkamp, G., Carlier, S.G., Mastik, F., Schon-

94

[35]

[36]

[37]

[38] [39] [40]

[41] [42] [43]

[44]

[45]

[46]

[47]

[48]

[49]

[50]

[51]

[52] [53] [54]

R.A. Baldewsing et al. / Intravascular Ultrasound Elastography eveld, A.H., Serruys, P.W., and Bom, N., Characterization of plaque components and vulnerability with intravascular ultrasound elastography, Phys Med Biol, Vol. 45, No. 6, pp. 1465–1475, 2000. Doyley, M.M., de Korte, C.L., Mastik, F., Carlier, S., and van der Steen, A.F.W. Advancing intravascular palpography towards clinical applications. Acoustical Imaging, Bristol, Plenum Press, New York, NY USA: 493–500; 2000. de Korte, C.L., Céspedes, E.I., van der Steen, A.F.W., and Lancée, C.T., Intravascular elasticity imaging using ultrasound: feasibility studies in phantoms, Ultrasound in Medicine and Biology, Vol. 23, No. 5, pp. 735–746, 1997. de Korte, C.L., Carlier, S.G., Mastik, F., Doyley, M.M., van der Steen, A.F.W., Céspedes, E.I., Serruys, P.W., and Bom, N. Intracoronary elastography in the catheterisation laboratory: preliminary patient results. IEEE Ultrasonics Symposium, Lake Tahoe, CA, USA: 1649–1652; 1999. Alam, S., Ophir, J., and Konofagou, E., An adaptive strain estimator for elastography, IEEE trans UFFC, Vol. 45, pp. 461–472, 1998. Brusseau, E., Perrey, C., Delachartre, P., Vogt, M., Vray, D., and Ermert, H., Axial strain imaging using a local estimation of the scaling factor from RF ultrasound signals., Ultras Imag, Vol. 22, pp. 95–107, 2000. Brusseau, E., Fromageau, J., Finet, G., Delachartre, P., and Vray, D., Axial strain imaging of intravascular data: results on polyvinyl alcohol cryogel phantoms and carotid artery, Ultrasound Med Biol, Vol. 27, No. 12, pp. 1631–42, 2001. Brusseau, E., Fromageau, J., Rognin, N.G., Delachartre, P., and Vray, D., Investigating elastic properties of soft biological tissues, IEEE Eng Med Biol Mag, Vol. 21, No. 4, pp. 86–94, 2002. Saijo, Y., Tanaka, A., Owada, N., Akino, Y., and Nitta, S., Tissue velocity imaging of coronary artery by rotating-type intravascular ultrasound, Ultrasonics, Vol. 42, No. 1–9, pp. 753–7, 2004. Perrey, C., Bojara, W., Holt, S., Lindstaedt, M., and Ermert, H., Intravascular ultrasound strain imaging with rotating single element transducer: initial in vivo experiments, Proceedings of the Second International Conference on the Ultrasonic Measurement and Imaging of Tissue Elasticity, p. 49, 2004. Perrey, C., Braeker, G., Bojara, W., Lindstaedt, M., Holt, S., and Ermert, H., Strain imaging with intravascular ultrasound array scanners: validation with phantom experiments, Biomed Tech (Berl), Vol. 48, No. 5, pp. 135–40, 2003. de Korte, C.L., Pasterkamp, G., van der Steen, A.F.W., Woutman, H.A., and Bom, N., Characterization of plaque components using intravascular ultrasound elastography in human femoral and coronary arteries in vitro, Circulation, Vol. 102, No. 6, pp. 617–623, 2000. Schaar, J.A., De Korte, C.L., Mastik, F., Strijder, C., Pasterkamp, G., Boersma, E., Serruys, P.W., and Van Der Steen, A.F.W., Characterizing vulnerable plaque features with intravascular elastography, Circulation, Vol. 108, No. 21, pp. 2636–41, 2003. de Korte, C.L., Sierevogel, M.J., Mastik, F., Strijder, C., Schaar, J.A., Velema, E., Pasterkamp, G., Serruys, P.W., and van der Steen, A.F.W., Identification of atherosclerotic plaque components with intravascular ultrasound elastography in vivo: a Yucatan pig study, Circulation, Vol. 105, No. 14, pp. 1627–1630, 2002. de Korte, C.L., Sierevogel, M., Mastik, F., Strijder, C., Velema, E., Pasterkamp, G., and van der Steen, A.F.W., Intravascular Elastography in Yucatan pigs: validation in vivo, European Heart Journal, Vol. 22, No. abst suppl, p. 251, 2001. de Korte, C.L., Carlier, S.G., Mastik, F., Doyley, M.M., van der Steen, A.F., Serruys, P.W., and Bom, N., Morphological and mechanical information of coronary arteries obtained with intravascular elastography; feasibility study in vivo, Eur Heart J, Vol. 23, No. 5, pp. 405–413, 2002. de Korte, C.L., Doyley, M.M., Carlier, S.G., Mastik, F., van der Steen, A.F.W., Serruys, P.W., and Bom, N. High resolution IVUS elastography in patients. IEEE Ultrasonics Symposium, Puerto Rico, USA: 1767– 1770; 2000. Schaar, J.A., Regar, E., Mastik, F., McFadden, E.P., Saia, F., Disco, C., de Korte, C.L., de Feyter, P.J., van der Steen, A.F.W., and Serruys, P.W., Incidence of vulnerable plaque patterns in humans: assessment with three-dimensional intravascular palpography and correlation with clinical presentation, Circulation, Vol. 109, in press, 2004. Virmani, R., Burke, A.P., Kolodgie, F.D., and Farb, A., Pathology of the thin-cap fibroatheroma: a type of vulnerable plaque, J Interv Cardiol, Vol. 16, No. 3, pp. 267–72, 2003. Schaar, J.A., Muller, J.E., Falk, E., Virmani, R., Fuster, V., Serruys, P.W., and others, Terminology for high-risk and vulnerable coronary artery plaques, Eur Heart J, Vol. 25, in press, 2004. Davies, M.J., Going from immutable to mutable atherosclerotic plaques, Am J Cardiol, Vol. 88, No. 4A, pp. 2F–9F, 2001.


95

[55] Baldewsing, R.A., De Korte, C.L., Schaar, J.A., Mastik, F., and Van Der Steen, A.F.W., Finite element modeling and intravascular ultrasound elastography of vulnerable plaques: parameter variation, Ultrasonics, Vol. 42, No. 1–9, pp. 723–9, 2004. [56] de Korte, C.L., Céspedes, E.I., van der Steen, A.F.W., and Lancée, C.T. Image artifacts in intravascular elastography. IEEE EMBS, Amsterdam, The Netherlands: paper no. 685; 1996. [57] de Korte, C.L., Céspedes, E.I., and van der Steen, A.F.W., Influence of catheter position on estimated strain in intravascular elastography, IEEE Trans Ultrason Ferroelectr Freq Control, Vol. 46, No. 3, pp. 616–625, 1999. [58] Ophir, J., Céspedes, E.I., Garra, B., Ponnekanti, H., Huang, Y., and Maklad, N., Elastography: ultrasonic imaging of tissue strain and elastic modulus in vivo, Eur J Ultrasound, Vol. 3, pp. 49–70, 1996. [59] Fung, Y.C. Biomechanics: mechanical properties of living tissues. New York, Springer-Verlag; 1993. [60] Gow, B.S., and Hadfield, C.D., The elasticity of canine and human coronary arteries with reference to postmortem changes, Circ Res, Vol. 45, No. 5, pp. 588–594, 1979. [61] Mozersky, D.J., Sumner, D.S., Hokanson, D.E., and Strandness, D.E., Transcutaneous measurement of the elastic properties of the humen femoral artery, Circulation, Vol. 46, pp. 948–955, 1972. [62] Salunke, N.V., and Topoleski, L.D., Biomechanics of atherosclerotic plaque, Crit Rev Biomed Eng, Vol. 25, No. 3, pp. 243–285, 1997. [63] Dulikravich, G.S., Martin, T.J., and Dennis, B.H. Multidisciplinary inverse problems. 3rd International Conference on Inverse Problems in Engineering, Port Ludlow, WA, USA; 1999. [64] Raghavan, K.R., and Yagle, A.E., Forward and inverse problems in elasticity imaging of soft tissues, Transactions on Nuclear Science, Vol. 41, No. 4, pp. 1639–1648, 1994. [65] Sumi, C., Suzuki, A., and Nakayama, K., Estimation of shear modulus distribution in soft tissue from strain distribution, IEEE Trans Biomed Eng, Vol. 42, No. 2, pp. 193–202, 1995. [66] Skovoroda, A.R., Emelianov, S.Y., and O’Donnell, M., Tissue elasticity reconstruction based on ultrasonic displacement and strain images, IEEE trans UFFC, Vol. 42, No. 4, pp. 747–765, 1995. [67] Kallel, F., and Bertrand, M., Tissue elasticity reconstruction using linear perturbation method, IEEE Transactions on Medical Imaging, Vol. 15, No. 3, pp. 299–313, 1996. [68] Sumi, C., and Nakayama, K., A robust numerical solution to reconstruct a globally relative shear modulus distribution from strain measurements, IEEE Trans Med Imaging, Vol. 17, No. 3, pp. 419–28, 1998. [69] Doyley, M.M., Meaney, P.M., and Bamber, J.C., Evaluation of an iterative reconstruction method for quantitative elastography, Phys Med Biol, Vol. 45, No. 6, pp. 1521–1540, 2000. [70] Bishop, J., Samani, A., Sciarretta, J., and Plewes, D.B., Two-dimensional MR elastography with linear inversion reconstruction: methodology and noise analysis, Phys Med Biol, Vol. 45, No. 8, pp. 2081–91, 2000. [71] Cohn, N.A., Kim, B.S., Erkamp, R.Q., Mooney, D.J., Emelianov, S.Y., Skovoroda, A.R., and O’Donnell, M., High-resolution elasticity imaging for tissue engineering, IEEE trans UFFC, Vol. 47, No. 4, pp. 956–966, 2000. [72] Zhu, Y., Hall, T.J., and Jiang, J., A finite-element approach for Young’s modulus reconstruction, IEEE Trans Med Imaging, Vol. 22, No. 7, pp. 890–901, 2003. [73] Van Houten, E.E., Paulsen, K.D., Miga, M.I., Kennedy, F.E., and Weaver, J.B., An overlapping subzone technique for MR-based elastic property reconstruction, Magn Reson Med, Vol. 42, No. 4, pp. 779–86, 1999. [74] Miga, M.I., A new approach to elastography using mutual information and finite elements, Phys Med Biol, Vol. 48, No. 4, pp. 467–80, 2003. [75] Moulton, M.J., Creswell, L.L., Actis, R.L., Myers, K.W., Vannier, M.W., Szabo, B.A., and Pasque, M.K., An inverse approach to determining myocardial material properties, J Biomech, Vol. 28, No. 8, pp. 935–948, 1995. [76] Beattie, D., Xu, C., Vito, R., Glagov, S., and Whang, M.C., Mechanical analysis of heterogeneous, atherosclerotic human aorta, J Biomech Eng, Vol. 120, No. 5, pp. 602–7, 1998. [77] Vorp, D.A., Rajagopal, K.R., Smolinski, P.J., and Borovetz, H.S., Identification of elastic properties of homogeneous, orthotropic vascular segments in distension, J Biomech, Vol. 28, No. 5, pp. 501–12, 1995. [78] Chandran, K.B., Mun, J.H., Choi, K.K., Chen, J.S., Hamilton, A., Nagaraj, A., and McPherson, D.D., A method for in-vivo analysis for regional arterial wall material property alterations with atherosclerosis: preliminary results, Med Eng Phys, Vol. 25, No. 4, pp. 289–98, 2003. [79] Soualmi, L., Bertrand, M., Mongrain, R., and Tardif, J.C. Forward and inverse problems in endovascular elastography. Acoustical Imaging. S. Lees and L. A. Ferrari. New York, USA, Plenum. 23: 203–209; 1997.

96


[80] Wan, M., Li, Y., Li, J., Cui, Y., and Zhou, X., Strain imaging and elasticity reconstruction of arteries based on intravascular ultrasound video images, IEEE Trans Biomed Eng, Vol. 48, No. 1, pp. 116–20, 2001. [81] Bank, A.J., Intravascular ultrasound studies of arterial elastic mechanics, Pathol Biol (Paris), Vol. 47, No. 7, pp. 731–7, 1999. [82] Kanai, H., Hasegawa, H., Ichiki, M., Tezuka, F., and Koiwa, Y., Elasticity imaging of atheroma with transcutaneous ultrasound: preliminary study, Circulation, Vol. 107, No. 24, pp. 3018–21, 2003. [83] Barbone, P.E., and Bamber, J.C., Quantitative elasticity imaging: what can and cannot be inferred from strain images, Phys Med Biol, Vol. 47, No. 12, pp. 2147–64, 2002. [84] Barbone, P.E., and Gokhale, N.H., Elastic modulus imaging: on the uniqueness and nonuniqueness of the elastography inverse problem in two dimensions, Inverse Problems, Vol. 20, pp. 283–296, 2004. [85] Baldewsing, R.A., Oomens, C.W.J., and van der Steen, A.F.W., Intravascular Young’s modulus reconstruction using a parametric finite element model., Proceedings of the 2003 IEEE International Ultrasonics Symposium, pp. 1879–1882, 2003. [86] Baldewsing, R.A., de Korte, C.L., Schaar, J.A., Mastik, F., and van der Steen, A.F.W., A finite element model for performing intravascular ultrasound elastography of human atherosclerotic coronary arteries, Ultrasound Med Biol, Vol. 30, No. 5, in press, 2004. [87] Baldewsing, R.A., Schaar, J.A., Mastik, F., Oomens, C.W.J., and van der Steen, A.F.W., IVUS modulography of vulnerable plaques using a parametric finite element model: validation on a phantom and human coronary artery, Proceedings of the Second International Conference on the Ultrasonic Measurement and Imaging of Tissue Elasticity, p. 75, 2003. [88] Baldewsing, R.A., Schaar, J.A., Mastik, F., Oomens, C.W.J., and van der Steen, A.F.W., Young’s modulus reconstruction for assessing vulnerable atherosclerotic plaque composition in vivo, Proceedings of the 2004 IEEE International Ultrasonics Symposium, 2004. [89] Schartl, M., Bocksch, W., Koschyk, D.H., Voelker, W., Karsch, K.R., Kreuzer, J., Hausmann, D., Beckmann, S., and Gross, M., Use of intravascular ultrasound to compare effects of different strategies of lipidlowering therapy on plaque volume and composition in patients with coronary artery disease, Circulation, Vol. 104, No. 4, pp. 387–92, 2001.


97

Computer Vision Analysis of Collagen Fiber Bundles in the Adventitia of Human Blood Vessels Pierre J. ELBISCHGER a,1 , Horst BISCHOF a , Gerhard A. HOLZAPFEL b and Peter REGITNIG c a Inst.

for Computer Graphics and Vision, Graz University of Technology, Austria for Structural Analysis, Graz University of Technology, Austria c Inst. of Pathology, Graz Medical University, Austria

b Inst.

Abstract. Numerical simulations, which are based on reliable biomechanical models of blood vessels, can help to get a better understanding of cardiovascular diseases such as atherosclerosis, and can be used to develop optimal medical treatment strategies. The adventitia is the outer most layer of blood vessels and its mechanical properties are essentially determined by the three-dimensional, structural arrangement of collagen fibre bundles embedded in the tissue. Global information such as the orientation statistics of the fibre bundles as well as detailed information as the crimp of the single fibres within the bundles is of particular interest in biomechanical modeling. In order to obtain a sufficiently large amount of data for biomechanical modeling, a fully automatic method for the structural analysis of the soft tissue is required. In this contribution we present methods based on computer vision to fulfill this task. We start by discussing proper tissue preparation and imaging techniques that have to be used to obtain data, which reliably represents the real three-dimensional tissue structure. The next step is concerned with algorithms that robustly segment the collagen fibre bundles and cope with various kinds of artifacts. Novel segmentation techniques for robust segmentation of individual fibril bundles and methods for estimation of their parameters, such as location, shape, mean fibril orientation, crimp of fibrils, etc., is discussed. The proposed algorithms are based on novel perceptual grouping methods operating on the extracted orientation data of fibrils. Finally, we demonstrate the results obtained by our fully automatic method on real data. In addition, for a more quantitative assessment, we introduce a generative structural model that enables the synthesis of three-dimensional fibre bundles with well-defined characteristics. Keywords. Biomechanic modeling, structural analysis, soft tissue, collagen fibers, microscopic images, computer vision

1 Corresponding Author: Pierre J. Elbischger, Inst. for Computer Graphics and Vision, Graz University of Technology, Inffeldgasse 16/II, 8010 Graz, Austria.

98

P.J. Elbischger et al. / Computer Vision Analysis of Collagen Fiber Bundles

1. Motivation and Background This book-chapter is concerned with an automatic method for structural analysis of collagen fibrils in biological soft tissues. This section motivates research both from a biomechanical point of view and a computer vision perspective. 1.1. Why do we Need to Understand the Concentration and Structural Arrangement, and the Structure-Function Relationships of Native Tissues? Soft tissues are wide-ranging biological materials in which the cells are separated by extracellular material. They may be distinguished from hard (mineralized) tissues such as bones for their high flexibility and their soft mechanical properties. Examples of soft tissues are tendons, ligaments, blood vessels such as arteries and veins, skins or articular cartilages among many others. Tendons are muscle-to-bone linkages that stabilize the bony skeleton (or produce motion), and ligaments are bone-to-bone linkages to restrict relative motion. Blood vessels are prominent organs composed of soft tissues which have to distend in response to pulse waves. The skin is the largest single organ (16% of the human adult weight). It supports internal organs and protects our body. Articular cartilages form the surface of body joints (which is a layer of connective tissue with a thickness of 1–5 mm), distribute loads across joints and minimize contact stresses, and friction (see [1] with more references therein). The mechanical behavior of soft biological tissues is strongly influenced by the concentration and structural arrangement of constituents such as collagen. The different behaviors of soft biological tissues commensurate with the different functions of individual tissues. A greater understanding of the foundations and interactions of structure and function of soft tissue, and, in particular, the associated mechanobiology is one of the research needs in the field. Mechanobiology is aimed to understand how cells change their structure and function in response to changes in their mechanical environment. Mechanobiology is a perfect complement of biomechanics each focuses on similar issues, just from different philosophic perspectives (for example, induction versus deduction) [2]. It also studies the mechanical factors that may be important in triggering the onset of atherosclerosis or aneurysms. Springer-Verlag has launched a new Journal in 2002 called Biomechanics and Modeling in Mechanobiology [3], which shows the importance of the field. A thorough understanding of the complex interrelation between mechanical factors and the associated biological responses may help to improve diagnostics which allow disease and injury to be treated earlier. It requires quantification of the mechanical environment of the involved tissues, i.e. geometrical and constitutive models of all tissue components involved. A greater understanding of the structure-function relationships of native tissues is also a prerequisite for appropriate repair and replacement tissues. Tissue engineering requires a clear understanding of the structural arrangements and functions of the associated living tissues, because much of gene expression is via mechanotransduction mechanisms. The success of tissue engineering is clearly based on knowledge of the biomechanics of native tissue. As a matter of fact, 1998 the United States National Committee on Biomechanics formed a subcommittee, which seeks to address challenges in repairing or replacing tissues that serve a predominantly biomechanical function. One essential goal of the subcommittee in advancing ‘tissue engineering’ is to ‘identify the


99

critical structural and mechanical requirements needed for each tissue engineered construct’. Hence, in particular, the quantitative knowledge of preferred orientations in biological soft tissues enhances the understanding of their general mechanical characteristics significantly. It is important to note that realistic structural models rely strongly on this knowledge. Collagen fibers are those components of many soft tissues that render the material properties anisotropic. In order to describe the anisotropic feature, appropriate geometrical data are required. They serve as an essential set of input data for numerical models. The focus of the proposed work will be on the concentration and structural arrangement of collagen in biological soft tissues, in particular, in arteries. 1.2. Background on the Structure of Soft Tissues – Collagen and Elastin Soft biological tissues of our body are complex fiber-reinforced composite structures. Their mechanical behavior is strongly influenced by constituents such as collagen and elastin, the hydrated matrix of proteoglycans, and the topographical site and respective function in the organism. Collagen. Collagen is a protein which is a major constituent of the extracellular matrix of soft tissue. It is the main load carrying element in a wide variety of soft tissues and is very important to human physiology (for example, the collagen content of (human) achilles tendon is about 20 times that of elastin). Collagen is a macromolecule with length of about 280 nm. Collagen molecules are linked to each other by covalent bonds building collagen fibrils. Depending on the primary function and the requirement of strength of the tissue the diameter of collagen fibrils varies (the order of magnitude is 1.5 nm [4]). The intramolecular crosslinks of collagen gives the soft tissues the strength which varies with age, pathology, etc. The function and integrity of organs are maintained by the tension in collagen fibers. In the structure of tendons and ligaments, for example, collagen appears as parallel oriented fibers [5], while many other tissues have an intricate disordered network of collagen fibers embedded in a gelatinous matrix of proteoglycans. As far as arteries are concerned collagen fibrils appear in the media (the middle layer of an artery). They are interconnected with muscle cells, and elastin within a complex three-dimensional network. The orientation of and close interconnection between the elastic and collagen fibrils, elastic laminae, and smooth muscle cells together constitute a continuous fibrous helix (Faserschraube) in the media [6,7]. The helix has a small pitch so that the fibrils in the media are almost circumferentially oriented. This structured arrangement gives the media high strength, resilience and the ability to resist loads in both the longitudinal and circumferential directions. From the mechanical perspective, the media is the most significant layer in a healthy artery. The adventitia is the outermost layer of the artery and consists also of collagen fibrils, which appear in thick bundles forming a fibrous tissue. In addition, fibroblasts and fibrocytes, and histological ground substance are present. The adventitia is surrounded continuously by loose connective tissue. The wavy collagen fibrils are arranged in helical structures and serve to reinforce the wall. They contribute significantly to the stability and strength of the arterial wall. The adventitia is much less stiff in the load-free configuration and at low pressures than the media. However, at higher levels of pressure

100


the collagen fibers reach their straightened lengths and the adventitia changes to a stiff ‘jacket-like’ tube which prevents the artery from overstretch and rupture. For a more detailed account of the structure (distribution and orientation) of the interrelated arterial components, the morphological structure and the overall functioning of blood vessels, see, for example, [8–11], Section 2, [1] and [12]. Elastin. Elastin, like collagen, is a protein which is a major constituent of the extracellular matrix of soft tissue. It is present as thin strands in soft tissues such as skin, lung, ligamenta flava of the spine and ligamentum nuchae (the elastin content of the latter is about five times that of collagen). The long flexible elastin molecules build up a three-dimensional (rubberlike) network, which may be stretched to about 2.5 of the initial length of the unloaded configuration. In contrast to collagen fibers, this network does not exhibit a pronounced hierarchical organization. As for collagen, 33% of the total amino acids of elastin consists of glycine. However, the proline and hydroxyproline contents are much lower than in collagen molecules. 1.3. Realistic Mathematical and Computational Models of Soft Tissues Rely Strongly on Histological Information A thorough understanding of the structural arrangement provides a foundation for functional integration across interacting biological processes, provide information for the design of powerful, and realistic mathematical and associated numerical models. It is wellaccepted and indisputable in the scientific community, that computer models and computational methods, when based on histological information, are necessary to understand the functions of living tissues and to gain more insights in the underlying mechanobiology. Of course, powerful mathematical and computational models are still the greatest needs in the field. Mathematical and computational models may predict observations across multiple length and time scales, and chemical, mechanical, and coupling responses of soft tissues. Computational methods are needed to model realistically many multidisciplinary processes encountered in biomechanics of soft tissues relevant to problems in medicine, surgery and bioengineering like diagnostic imaging, surgical planning and intervention, medical therapy, and biomedical engineering design for tissue engineering or medical devices. They are needed to solve the complex geometries and loading conditions, and the initial and boundary-value problems of clinical, industrial, and academic importance. Computational models may (i) predict the risk of rupture of aneurysms, (ii) identify the failure strength of an anterior cruciate ligament, (iii) improve diagnostics and therapeutical procedures that are based on mechanical treatments, (iv) study the mechanical factors that may be important in triggering the onset of aneurysms or atherosclerosis, the major cause of human mortality in the western world, (v) describe the pulse wave dynamics and the interaction between the heart and the circulatory system, (vi) investigate the changes in the arterial system due to age, disease, hypertension and atherosclerosis, which is of fundamental clinical relevance.


101

1.4. The Focus of the Chapter Quantitative data about the structural arrangement of collagen are particularly rare for single layers of human arteries (i.e. intima, media and adventitia). In order to develop reliable biomechanical models, it is crucial to have a fast, reliable and inexpensive tool available for the identification of the structural arrangement of collagen fiber bundles in the tissue. To relieve medical assistants from the tedious work of measuring structural data by hand, which is also prone to errors, it is reasonable to strive for an analysis of tissues by means of computer vision methods. The use of computer vision methods for the extraction of relevant structural data guarantees an objective (robust against human subjective impressions) way for data acquisition, and, furthermore, allows an exhaustive, automatic analysis of a tissue sample set, large enough to gain statistically meaningful results. The focus of this chapter is the identification of the structural arrangement of collagen in biological soft tissues, in particular, in the outermost layer of human arteries, i.e. the adventitia (tunica externa). Our analysis is based on transmitted light microscopy (TLM) images of histological tissue samples. Once, the automatic structure analysis problem of adventitial tissue is solved, the algorithms can easily be adopted for similar problems, e.g. the structure analysis of the intima and media. Compared to other imaging techniques such as magnetic resonance imaging (MRI), computer thomography (CT) or microtomography (Micro-CT), where in particular the contrast of collagen fibers almost vanishes, TLM images of stained soft tissues feature an extremely high contrast. This fact makes an automatic computer vision-based analysis procedure possible. Images from TLM do not contain already coded orientation information such as images obtained by polarized light microscopy (PLM), quantitative polarized light microscopy (QPLM) or small angle light scattering (SALS), however, they do represent the full structure information in a single image and additionally provide the possibility to segment artifacts. Our intention and goal is to extract and group orientation information of collagen fibers by means of modern computer vision techniques and end up with a powerful approach that is capable of providing detailed structure information about the specimen. The major benefits of our approach are: (i) with a resolution of 0.25 μm our setup is capable of resolving individual collagen fibers. (ii) The TLM setup is inexpensive and easy to access for any scientist. (iii) The required histological preparation techniques are inexpensive and well established. (iv) The problem of structural information extraction is entirely solved in software and does not require any expensive equipment in the imaging setup. (v) The analysis is performed fully automatic, except the image capturing process. 1.5. Related Work While a great deal of research effort in the field of medical computer vision has been devoted to the analysis of MRI, CT, Micro-CT and ultra sound images, there is little research performed on the automatic analysis of light microscopic histological images. Most studies that address the extraction of structural data related to histological microscopic images focus on blobs and cell shaped structures that represent blood cells or cell nucleus rather than fibrous structures and their precise structural information (see, for example, [14–18]). Because of the importance of collagen fibers, many researchers in the areas of medicine, biology and biomechanics have tried to reveal its structural information such as ori-

102


entation, density, coherence, etc. using various techniques. Most of these studies do not use automatic algorithms for segmenting the structure in question. In the following we give a brief overview about the recent research in the field of fibrous structure analysis. High directional correlations between the long axes of cellular shapes and nuclei of, for example, a stained patch of a media may be studied under a microscope. The very first contribution concerning the directional analysis of smooth muscle cells seems to be attributed to [19]. At a later date, [20] used a combination of microdissection and etching technique to show that dissected threads of muscle tissues are helically shaped with constant radius. The author concluded that the helical patterns are also present in muscular arteries and that the media is composed mainly of layered structures of helices. Quantitative studies regarding (helical) pitch angles were carried out by [21–23] and [24] for several organisms at different locations. These studies do not support the conclusion of a unified orientation of smooth muscle cells drawn by [20]. The non-uniformity of fiber orientations in specimens originating from different organisms and different locations was pointed out by [25]. Smooth muscle cells provide active contractile elements of arterial walls and have long and thin centrally located nuclei (see [22] and [26]), which can be stained with hematoxylin in order to reveal the cell (nuclei) orientations ( [25]). A number of techniques have been proposed for analyzing the orientation of nuclei: [25] used a digitizer to enter manually the coordinates of the nuclei of the human intracranial media into a computer. Thereby, the orientation is described by two angles (in the histological plane and in an orthogonal plane). In addition, [27] used a graphic tablet to digitize the coordinates of nuclei within the human brain artery. As a reference frame of coordinate axes they embedded a precisely trimmed block together with the vessel segment in paraffin wax. The most thorough study was undertaken by [28], in which shape, position, linear dimensions, volumes and orientation of nuclei within several vessel types of male Wistar rats are determined by means of computer-aided analyses. Sections with 0.5 μm thickness are produced by an ultramicrotome and then stained and cut again to produce ultra-thin sections with 60–80 (nm) thickness. Hence, the digitized contours of cells are assembled into a 3D model by appropriate software tools. Recent studies concentrate on SALS [29–32] and on PLM [32,33] using the birefringent optical property of collagen fibers in order to obtain information about their structural arrangement [34]. PLM is generally recommended as the most appropriate method for detecting, describing and interpreting wave-like structures [35,36], whereas SALS makes it possible to detect the fiber orientation. SALS provides a maximal spatial resolution of about 50 μm and, therefore, prevents a detailed analysis of individual fibers that requires a spatial resolution of at least 0.5 μm. Traditional crossed PLM provides essentially qualitative information (they detect birefringent materials) rather than quantitative measurements of the spatial distribution of the optical anisotropy of the specimen. So called QPLM [37] is capable to overcome this shortcoming of crossed PLM. Tower et al. [38] proposed a method to analyze the collagen fiber alignment during mechanical testing of soft tissues based on QPLM. With respect to our approach the technique proposed in [38] uses much thicker samples and, therefore, integrates the optical signal across the whole sample. The retardation does not distinguish between alignment among multiple fiber populations (for example, collagen and elastic fibers) nor does it account for layers with distinct alignment fields. The power of the method is, that the structural analysis can be performed dynamically during mechanical testing of soft tissues. Mas-


103

soumian et al. [39] describe a modification to the optical system of a confocal microscope which analyzes the polarization state of the image light and came up with an enhanced method for QPLM. The system uses a novel form of rotating analyzer which, together with lock-in detection, permits images to be obtained where the image contrast corresponds to both specimen retardance and orientation. Compared to our approach which is based on histological samples, the special equipment required for a QPLM setup is much more complex and expensive. Most recently laser scanning microscopy (LSM) experiences an increasing interest for the analysis of collagen fibers. In contrast to the above mentioned techniques, LSM is non-destructive and produces volume data directly. Axer et al. [40,41] performed a coarse analysis of the 3D architecture of the collagen fibers in linea alba and rectus sheaths. The analysis was done manually. By using a commercial software, Young et al. [42] registered several LSM volume images to each other and performed a structural analysis of myocardial laminae and collagen network in rat hearts by visual inspection. The preliminary study [43] is aimed to investigate small angle X-ray scattering (SAXS) diffraction patterns of isolated arterial layers during tensile testing. The application of this analyzing technique to identify the structure of human soft tissues is new. It is a challenging approach for the analysis of very small periodic structures, however, it may not be appropriate for a precise macroscopic analysis. The diffraction pattern obtained in the preliminary study [43] is highly specific for collagen fibrils. Highest intensities were observed for the adventitia, followed by the intima. Medial layers showed predominately diffuse scatter and only a weak first order maximum. In particular, for the adventitia and the intima SAXS diffraction data in combination with tensile testing may provide valuable data for micro-and nano-structural constitutive modeling. Most of the above mentioned studies do not make extensively use of image processing techniques to automatically extract structural data. They either require massive manual interaction or use the output of an expensive analyzing setup directly (as, e.g., for QPLM). In the following we address approaches that make use of image processing algorithms. Smooth muscle cells provide active contractile elements of arterial walls and have long and thin centrally located nuclei (see [22] and [26]), which can be stained with hematoxylin in order to reveal the cell (nuclei) orientations ( [25]). A number of techniques have been proposed for manually (computer-aided) analyzing the orientation of nuclei [25,27,28]. In a recent paper [11] an automatic technique was postulated for obtaining information about preferred orientations in isolated arterial tissues, and, additionally, the concentration of nuclei. The authors assumed that the orientation of the muscle cells correlates with the orientation of the collagen fibers, which enabled them to estimate the orientation statistics by analyzing the muscle cell nuclei. For the adventitia, however, the concentration of muscle cells is very small so that the intra-spatial voids between collagen fiber bundles may be used as indicators for preferred orientations of the tissues. An interesting 2D image analysis technique to characterize the collagen morphology of articular cartilage, which is the most similar to our approach, was proposed by Xia and Elder [44]. They digitized transmission electron microscopy (TEM) images and divided them into small square ROIs. A region was chosen to be large enough to include several fibrils. The morphology of each small and localized region, was characterized by three quantities: the concentration of the fibrils, the overall orientation of the fibrils, and the anisotropy of the fibrils. Nevertheless, they neither perform any grouping operation,

104


which is essential in order to obtain reliable orientation information, nor any type of artifact segmentation to reject regions that hold no meaningful information. Furthermore, they require an expensive setup and need to digitize analog film images before the image processing can take place. Wu et al. [45] demonstrate a vision algorithm designed to extract quantitative structural information about individual collagen fibers (orientation, length and diameter) from LSM images of collagen gels. They trace individual fibers in the volume data set. Althoug the LSM seems to be a challenging approach it has some drawbacks. It is expensive and, therefore, restricted to a small group of researchers. Using a stack of registered histological cross-section images, our approach can be extended to 3D too. Because LSM is based on back scattered light from single fibers, the maximal analyzing depth for our samples is restricted to approximately 60 μm and the contrast of collagen as well as the spatial resolution decrease with increasing depth. Glazer [46] proposed an algorithm to segment single and well separative fibrils in TLM images. However, the proposed algorithm is too weak to extract the complex, structural and hierarchical data of collagen fibers from the microscopic images of soft tissues. It neither works on texture-like images nor does it perform any kind of fibril bundle grouping. Beside collagen fiber analysis, Eberhardt and Clarke [47] used X-ray to nondestructively probe the internal structure of several fibrous materials (glass fibre reinforced composites and non-woven textile samples) in 3D. They are also the authors of a book about microscopy techniques in material science [48], in which they address an image processing based automatic analyzes of fibrous structures. In their analyzes they use the profile of single fibers (cycles and ellipses), which are larger than 30 μm in diameter, to obtain structural information. Because the diameters of collagen fibers (1 μm) are far smaller, these methods cannot be applied.

2. Imaging Techniques for Collagen A reliable structure analysis of collagen require a proper data set (images) that comprehensively represent the structural information of collagenous components within soft tissues. The properties and the quality of the raw data set strongly determine the accuracy of the entire analysis process. Hence, a careful choice of the used imaging technique is required. Important issues in the image acquisition process are: • Keep the number of artifacts as small as possible. • High contrast of the collagenous structures compared with artifacts and other tissue components (elastic fibers, muscle tissue, etc.). • Choose an ‘optimal’ image resolution that provides adequate and detailed information (fibrils) as well as global information (fibrous bundles). • The data base should enable extraction of 3D data. • Keep the data volume as small as possible. • Establish an easy and reproducible setup for soft tissue analysis. In order to analyze collagen structures, different methods have been investigated and assessed. To avoid autolysis of the tissue, the time period between death and tissue fixation (use, e.g., 4% formalin, 70% alcohol or bouin) is of particular importance and should be as close to the mortal time as possible.


105

In the assessment of the different imaging techniques the following issues are particularly of interest. Destructiveness: Is the method non-destructive or destructive? If the tissue has to be sliced the method is destructive and a registration of successive slices will be necessary in order to reveal the 3D structure. The registration process may be difficult as a result of the non-rigid tissue deformations during the preparation process, see [49]. On the other hand only one non-destructive imaging technique that achieves the necessary resolution is currently available, namely the laser scanning microscopy (LSM). Resolution in x and y-directions: A single collagen fiber has a thickness of about 1–2 μm, which requires the imaging technique to have a minimal resolution of 500 nm. Resolution in z-direction: Some imaging techniques require a mechanical slicing of the specimen into a stack of thin slices and others provide the possibility to scan an entire tissue volume at once. In both cases the slices that build up the volume have to be small enough to pursue the fiber bundles through the volume. Because entire fiber bundles and not single fibers should be pursued through the image stack, the spatial resolution in z-direction may be lower than in x and y-directions. For TLM it has been shown, that ∼ 3 μm thick slices provide sufficient information. The occlusion of the single fibers at a slice thickness of ∼ 3 μm is moderate and allows the estimation of a single fiber and, therefore, the segmentation of a fiber bundle in 2D. Contrast of the structures in question: The way how the different structures within the tissue appear in the image, influence the reliability of the segmentation of collagen. If all structures have a good contrast it will be easier to obtain a correct segmentation. Fixation of the tissue: Typically the fixation causes a shrinkage of the tissue and, therefore, alters the structure of the original tissue. In order to keep this influence small, the correct fixation has to be chosen. Typical values of contractions for different fixations are: formalin (5%–10%), alkohol (10%), bouin (99%).24 Therefore, a calcium score of 0, denoting no evidence of coronary calcium, can virtually exclude those patients with obstructive CAD, making this test an effective screen prior to invasive angiography. An important point in the interpretation of CAC scores relates to the detection of obstructive CAD. A negative test, no evidence of calcified atherosclerotic plaque, can virtually exclude obstructive disease. A positive EBCT study, presence of CAC, is nearly 100% specific for atheromatous coronary plaque.25 However, since both obstructive and non-obstructive lesions have calcification present, CAC is not specific to obstructive disease.26 While increasing calcium scores are more predictive of obstructive CAD, there is not a 1:1 relationship between calcification and stenosis. The overall specificity of any

J.S. Shinbane and M.J. Budoff / Computed Tomographic Cardiovascular Imaging

151

CAC for obstructive CAD is approximately 66%.27 In a study of 1851 patients undergoing angiography and CAC measure,28 EBCT calcium scanning in conjunction with pretest probability of disease derived by a combination of age, gender and risk factors, could assist the clinician in predicting the severity and extent of angiographically significant CAD in symptomatic patients. EBCT is comparable to nuclear exercise testing in the detection of obstructive CAD.26,29 As opposed to stress testing, the accuracy of EBCT is not limited by concurrent medications, ability to exercise, or baseline electrocardiogram abnormalities. Role of Coronary Calcium in Risk Stratification Disease processes related to atherosclerosis are the primary cause of morbidity and mortality in industrialized nations.30 Identification of individuals at high risk for cardiovascular events is important to primary prevention strategies, as lifestyle modification and pharmacologic therapies can impact life expectancy in high-risk persons.31 The initial manifestation of CAD is a myocardial infarction (MI) or death in up to 50% of patients.32 Most cardiac events occur in the intermediate risk population, where aggressive riskfactor modification is often not often recommended or applied. Unfortunately, traditional risk factor assessment helps predict only 60–65% of cardiac risk, therefore many individuals without established risk factors for atherosclerotic heart disease continue to experience cardiac events.33 An understanding of coronary plaque pathophysiology is necessary to recognize some of the limitations of traditional cardiovascular risk stratification. Arterial segments with obstructive coronary plaque (stenosis in the artery of >50% severity) are often not the site of vessel occlusion in acute MI.34−37 Acute coronary occlusion most frequently occurs at the site of mild to moderate stenoses in association with the process of plaque rupture. Therefore, plaque burden, and not stenosis severity, is a more important marker of disease. Tests assessing for evidence of obstructive CAD, such as exercise testing or pharmacologic nuclear or echo cardiac imaging, will not identify a significant number of asymptomatic patients with atherosclerotic plaque who are at risk for acute MI. Studies of patients dying from either acute MI or sudden cardiac death have demonstrated that the extent of coronary atherosclerosis, rather than the severity of stenosis, is the most important predictor.38 These studies emphasize the importance of measurement of atherosclerosis burden in the assessment of risk for future cardiovascular events, rather than relying solely on evidence of obstructive coronary artery disease. Coronary Artery Calcium in Symptomatic Individuals Studies have demonstrated that CAC has prognostic significance in symptomatic individuals.39−43 Margolis et al.39 assessed the significance of CAC found on fluoroscopy in 800 patients undergoing coronary angiography. The patients with CAC had a 5-year survival rate of 58%, compared with a rate of 87% for the patients without CAC. A study of 192 patients observed for an average of 50 +/− 10 months, after undergoing an EBCT study while in the emergency department for chest discomfort, found that the presence of CAC (calcium score >0), and increasing absolute calcium score values were strongly related to the occurrence of hard events (p < 0.001) and all cardiovascular events (p < 0.001).43 The patients with absolute calcium scores in the top 2 quartiles

152


had a relative risk of 13.1 (95% CI, 5.6 to 36; p < 0.001) for new cardiovascular events as compared with the patients in the bottom 2 quartiles. The annualized cardiovascular event rate was 0.6% for subjects with a coronary artery calcium score of 0 compared with an annual rate of 13.9% for patients with a coronary artery calcium score >400 (p < 0.001). Additional studies have shown not only that CAC provides prognostic information, but that the extent of CAC provides more prognostic information than angiography or risk factors in symptomatic patients. A multicenter study of 491 patients undergoing coronary angiography and EBCT scanning found that higher calcium scores were associated with an increased risk of coronary events over the next 30 months as compared to patients in the lowest quartile of score.41 In multivariate analysis, the only predictor of a hard cardiac event was log calcium score. In another study of symptomatic patients, EBCT detected CAC was a stronger independent predictor of disease and future events than a sum of all of the traditional risk factors combined.40 Keelan et al.42 followed 288 symptomatic persons who underwent angiography and EBT calcium scanning for a mean of 6.9 years, and found age and CAC score were the only independent predictors of future hard coronary events. Coronary Artery Calcium in Asymptomatic Individuals CAC is also a useful predictor of cardiovascular events in asymptomatic individuals. Unfortunately, at least half of all first coronary events occur in asymptomatic individuals who are unaware that they have developed CAD. Often the initial clinical presentation is sudden death or acute myocardial infarction (MI).32 Several lipid-lowering trials have shown that significant risk reduction can be attained with both secondary and primary prevention measures.31,44 However, these studies have required study of large patient populations to demonstrate benefit, and documented that risk reduction is only on the order of 25–37%, suggesting that more effective screening modalities are needed to identify asymptomatic individuals at risk of cardiac events. Several prospective trials have demonstrated the prognostic ability of EBCT to identify asymptomatic patients at high risk of cardiac events. Arad et al.45 initially reported a 19 month follow-up of 1,173 patients. Asymptomatic individuals were scanned using EBCT as well as measures of traditional risk factors, and followed prospectively for cardiac events. This study demonstrated CAC to be the strongest predictor of future cardiac events, with patients in the highest score category over 20 times more likely to suffer a cardiac event (odds ratio 22.3, CI 5.1–97.4). This prospective study now has been carried out for a total of 3.6 years of follow-up, maintaining the strong power of this technology to predict future cardiac events.46 The subjects who had events had significantly higher calcium scores than did the subjects with no events (764 +/− 935 vs 135 +/− 432, p < 0.0001). A calcium score ≥160 was associated with a high likelihood of having a soft event (odds ratio, 15.8) or a hard event (odds ratio, 20.2). The predictive ability of the absolute calcium score was excellent for all coronary events and for hard events alone. Detrano et al.47 published an analysis in an older population (1196 patients, 89% male, mean age 66 years). This study demonstrated that while CAC was a significant predictor of future cardiac events, short term follow-up demonstrated it did not have great power over traditional risk factors to discriminate who will develop CAD events. This study utilized scanning protocols (6 mm thick slices) that have since proven to be not


153

as sensitive as the standard protocol (3 mm slices), and a definition of a calcific focus that was up to 16 times greater than other studies (>8 mm2 ).48 Long-term (median 7.0 years) follow-up has demonstrated that the CAC score adds predictive power beyond that of standard coronary risk factors Subsequent data in this same cohort49 have demonstrated that EBCT coronary calcium scores were incremental in predicting cardiac risk, consistent with prior studies. Greenland et al.52 published longer term follow up of the South Bay heart watch cohort. CAC was found to be predictive of risk in patients with a Framingham Risk Score of >10%, with a high CAC score able to predict risk beyond Framingham risk score alone. As compared to a CAC score of 0, a CAC score of >300 was highly predictive of cardiac events (HR 3.9, p QCCA (PCMR). This difference can be made smaller by different choices of boundary conditions, or by using compliant vessel wall models. 9.4. Prediction of Peak Velocity A model of the left carotid artery of a patient with atherosclerosis was reconstructed from contrast-enhanced MRA images. The MIP of the MRA images and the reconstructed model are shown in Fig. 8. In this case, the patient had an occluded right carotid artery and a normal left carotid artery. Flow rates in the CCA, ICA and ECA were measured with phase-contrast MR and the peak velocity in the bulb region was measured with Doppler ultrasound (DUS). A CFD simulation was performed imposing flow rates measured with PC-MR in the CCA and ICA. Traction-free boundary conditions were prescribed in the ECA. Although the PC-MR images were quite noisy, the CFD and PC-MR velocity profiles in the region of the bifurcation are in good agreement (see Fig. 8). A comparison of the peak velocity in the bulb region obtained with the CFD model and measured with DUS is shown in Fig. 8. The DUS velocity waveform was obtained from the envolvent of the ultrasound spectrum and averaged over several cardiac cycles. Figure 8 shows a very good agreement between the CFD and DUS peak velocities.

P. Yim et al. / Characterization of Shear Stress on the Wall of the Carotid Artery

433

Figure 8. Validation of peak velocity and profiles.

Figure 9. Reconstructed anatomical models.

10. Examples In this section we present examples of finite element models of carotid artery hemodynamics in patients with different degrees of atherosclerotic disease. Patient-specific models were constructed for each of these arteries as described before. The reconstructed vascular models are shown in Fig. 9. These models correspond to carotid arteries with the following degrees of atherosclerotic disease: a) normal, b) nor-

434


Figure 10. Details of finite element grids.

Table 1. Summary of hemodynamic calculations. No.

p (mmHg)

peak mean WSS (dyne/cm2 )

Disease

a

12.0

50.0

Normal

b

2.0

100.0

Normal - contralateral occlusion

c

5.0

40.0

Mild

d

10.0

30.0

Moderate

e

18.0

200.0

Severe

f

20.0

200.0

Severe - kink

mal with occluded contralateral internal carotid artery, c) mild stenosis, d) moderate stenosis, e) severe stenosis, f) severe stenosis. Tubular deformable models followed by surface merging were used in the first three cases (a,b,c) and iso-surface deformable models were used in the last three cases (d,e,f). The finite element grids generated for each of these models are shown in Fig. 10. Blood flow analyses were conducted in each of these models for two cardiac cycles. In a post-processing stage, hemodynamic quantities such as peak pressure drop (pressure difference between the CCA inlet and the ICA outlet at peak systole) and mean or timeaveraged wall shear stress (WSS) magnitude were calculated. The results are summarized in Table 1. Visualizations of the computed peak pressure drops, are presented in Fig. 11. A slight pressure increase can be observed at the apex of the bifurcation and increased pressure drops across stenoses can be clearly seen. Visualizations of the distribution of mean WSS are shown in Fig. 12. Regions with increased WSS include, arterial stenoses, the apex of the carotid bifurcation, and outer surface of vessels with high curvature. Low WSS are observed in the bulb region of healthy carotid arteries. The oscillatory shear index (OSI) provides a measure of the angular change of the shear force during the cardiac cycle [56]: || 1 1− (59) OSI = 2 < |f| >


435

Figure 11. Peak pressure distributions.

where f = τ · n is the tangential force per unit area, n is the normal to the vessel surface, and denotes time average over the cardiac cycle. Regions with elevated values of OSI indicate that the direction of the shear force changes significantly over the cardiac cycle. These changes are thought to possibly lead to endothelial cell damage and development or progression of the atherosclerotic disease. However, one must be cautious because the OSI does not account for the magnitude of the shear force, in other words, the shear force may be changing in direction significantly (high OSI), but its magnitude may be very low. Visualizations of the OSI distribution (Fig. 13) indicate that high levels of OSI are encountered near the bulb, where the flow recirculates, or at the bifurcation apex, where the flow divides. More complex distributions of OSI observed in stenosed vessels may be indicative of disturbed or turbulent flow patterns. The effects of the non-Newtonian properties of blood are illustrated in Fig. 14. A model of a healthy carotid artery was analyzed both with a Newtonian approximation and the Casson model. As shown in the figure, differences of the order of 10% to 20% in the distribution of the mean wall shear stress magnitude can be observed in the region of the bulb. Further investigation is needed to determine the importance of nonNewtonian effects for varying degrees of atherosclerotic disease. Incorporating non-

436


Figure 12. Mean wall shear stress distributions.

Newtonian models into a CFD solver is very simple and the computing time is only slightly affected (it increases by about 10%). In addition, estimating a priori the regions where non-Newtonian effects may be important is difficult because it requires knowledge of the time varying flowfield, and in particular the regions of low flow and low shear. For these reasons, we favor the use of non-Newtonian models in all calculations. When using such models, one must be careful with the initial conditions used to initialize the flowfield. If the flow starts from zero, the initial non-Newtonian viscosity would be maximum, and it may take several cardiac cycles to adjust to the correct values. Alternatively, the flowfield can be initialized from a simulation performed using a Newtonian model for an entire cardiac cycle or from a steady run using the initial values of the flow rates. Finally, the complex flow patterns computed for patients with carotid artery disease, are visualized in Fig. 15. These visualizations were produced by cutting the computational domain along a surface that follows the axis of each arterial branch [86]. Velocity magnitudes on these cuts were plotted at five instants during the cardiac cycle. Several important flow characteristics can be observed. These include: high speed jets at arter-


437

Figure 13. Distributions of oscillatory shear index.

Figure 14. Non-Newtonian effects on the wall shear stress distribution.

ial stenoses, flow recirculation zones in the bulb, highly skewed velocity profiles in regions of high curvature, and unstable velocity patterns distal of severe stenoses. Note that flow impingement correspond with regions of increased pressure, and regions of highly skewed velocity profiles with increased wall shear stress.

438


Figure 15. Visualization of flow patterns for cases with atherosclerotic disease.

11. Discussion Progress in characterization of hemodynamic conditions in the carotid artery has been substantial. Namely, a comprehensive computational process has been assembled to perform vascular CFD analysis. The reliability of that methodology has been established, on a preliminary basis, in a diverse set of experiments. However, several legitimate concerns must still be addressed before vascular CFD analysis can be used on a routine basis and without significant technical expertise. A primary concern is the high computational complexity (computing time) of vascular CFD analysis. The computational schemes described in this chapter for obtaining the finite-element solution are specifically geared towards reduction of the computational complexity. These schemes now permit the computation of the finite-element solution of the carotid artery hemodynamics in less than 12 hours on a consumer PC that is a remarkable performance considering that such analysis was only recently relegated to supercomputers. However, such the delay currently in-


439

volved in performing vascular CFD analysis is probably unacceptable in a clinical setting. Perhaps of even greater concern is that vascular CFD analysis requires expert supervision, for example, for ensuring that the correct segmentation of the vessel has been obtained and for identifying and truncating inflow and outflow locations. Of the greatest concern still, is whether the CFD analysis can accurately assess the hemodynamic conditions in the carotid artery in a consistent manner. The accuracy of vascular CFD analysis might be compromised by any of a variety of factors including inaccuracy of the vessel segmentation, inaccuracy of the flow quantification and the inaccuracy of the simplifying assumptions of the fluid-dynamics modeling. In spite of these technical concerns, it is now appropriate to begin consideration for the use of vascular CFD analysis in the clinical setting. Ultimately, CFD analysis could take its place, along with other imaging methods, in the evaluation of carotid artery disease.

References [1] Petty, G.W., Brown, R.D., Whisnant, J.P., Sicks, J.D., O’Fallon, W.M., Wiebers, D.O., Ischemic stroke subtypes; a population-based study of functional outcome, survival, and recurrence, Stroke, vol. 31, pp. 1062–1068, 2000. [2] Barnett, H.J.M., Taylor, D.W., Eliasziw, M. et al., Benefit of carotid endarterectomy in patients with symptomatic moderate or severe stenosis, N Engl J Med, vol.339, pp. 1415–25,1998. [3] NASCET, Beneficial effect of carotid endarterectomy in symptomatic patients with high-grade carotid stenoses, New England Journal of Medicine, vol.325, pp. 445–53, 1991. [4] Constantinides, P. Plaque fissures in human coronary thrombosis, J Atheroscler Res, vol.6, pp. 1– 17,1966. [5] Farb, A., Burke, A.P., Tang, A.L., Smialek, J., Virmani, R. Coronary plaque erosion without rupture into a lipid core, Circulation, vol. 93, pp. 1354–1363, 1996. [6] Arbutini, E., Dal Bello, B., Morbini, P., Burke, A.P., Bocciarelli, M., Specchia, G., Virmani,R., Plaque erosion is a major substrate for coronary thrombosis in acute myocardial infarction, vol. 82, pp. 269– 272,1999. [7] Fernandez-Ortiz, A., Falk, E., Fuster, V., Meyer, B., et al., Characterization of the relative thrombogenicity of atherosclerotic plaque components: implications for consequences of plaque rupture, J Am Coll Cardiol, vol.23, pp. 1562–9, 1994. [8] Cheng, G.C., Loree, H.M., Kamm, R.D., Fishbein, M.C., Lee, R.T., Distribution of circumferential stress in ruptured and stable atherosclerotic lesions. A structural analysis with histopathological correlation, Circulation, vol.87, pp. 1179–87,1993. [9] Arroyo, L.H., Lee, R.T., Mechanisms of plaque rupture: mechanical and biologic interactions, Cardiovascular Research, vol.41, pp. 369–375,1999. [10] Falk, E., Why do plaques rupture?, Circulation, vol. 86III, pp. 30–42, 1992. [11] Park, A.E., McCarthy, W.J., Pearce, W.H., Koushanpour, E., Pearce, W.H., Yao, J.S., Carotid plaque morphology correlates with presenting symptomatology, J Vasc Surg, vol. 27, pp. 872–9,1998. [12] Bluestein, D., Gutierrez, C., Londono, M., Schoephoerster, R.T., Vortex Shedding in steady flow through a model of an arterial stenosis and its relevance to mural platelet deposition, Annals of Biomedical Engineering, vol. 27, pp. 763–773,1999. [13] Phelps, J.E., DePaola, N., Spatial variations in endothelial barrier function in disturbed flows in vitro, Am J Physiol, vol. 278, pp. H469–H476, 2000. [14] DePaola, N., Gimbrone, M.A., Davies, P.F., Dewey, C.F., Vascular endothelium responds to fluid shear stress gradients, Arteriosclerosis and Thrombosis, vol. 12, pp. 1254–1257, 1992. [15] Chappell, D.C, Varner, S.E, Nerem R.M., et al., “Oscillatory shear stress stimulates adhesion molecule expression in cultured human endothelium,” Circ Res, vol. 82, pp. 532–9, 1998. [16] Dewey, C.F., Bussolari, S.R., Gimbrone, M.A., Davies, P.F., The dynamic response of vascular endothelial cells to fluid shear stress, J Biomech Eng, vol. 103, pp. 177, 1981. [17] Burrig, K.F., The endothelium of advanced arteriosclerotic plaques in humans, Arteriosclerosis and Thrombosis, vol. 11, pp. 1678–1689, 1991.

440


[18] Burrig, K.F., The endothelium of advanced arteriosclerotic plaques in humans, Arteriosclerosis and Thrombosis, vol. 11, pp. 1678–1689, 1991. [19] Dimmeler, S., Hermann, C., Zeiher, A.M., Apoptosis of endothelial cells. Contribution to the pathophysiology of atherosclerosis?, Eur Cytokine Netw. vol. 9, pp. 697–8, 1998. [20] Norris, J.W., Changing perspectives in asymptomatic carotid stenosis, in Stroke; Pathophysiology, Diagnosis, and Management, HJM Barnett, JP Mohr, BM Stein, FM Yatsu eds, Churchill Livingstone, 1992. [21] Blankensteijn, J.D., van der Grond J., Mali, W.P., Eikelboom, B.C., Flow volume changes in the major cerebral arteries before and after carotid endarterectomy: an MR angiography study, Eur J Vasc Endovasc Surg, vol. 14, pp. 446–50, 1997. [22] Ledru, F., Theroux, P., Lesperance, J., et al., Geometric features of coronary artery lesions favoring acute occlusion and myocardial infarction: a quantitative angiographic study, J Am Coll Cardiol, vol. 33, pp. 1353–61, 1999. [23] Rothwell, P.M., Warlow, C.P., Low risk of ischemic stroke in patients with reduced internal carotid artery lumen diameter distal to severe symptomatic carotid stenosis: cerebral protection due to low poststenotic flow? On behalf of the European Carotid Surgery Trialists’ Collaborative Group. Stroke., vol. 31, pp. 622–30, 2000. [24] Bech, G.J., De Bruyne, B., Akasaka, T., Liistro, F., Bonnier, H.J., Koolen,J.J., Pijls, N.H., Coronary pressure and FFR predict long-term outcome after PTCA. Int J Cardiovasc Intervent., vol. 4, pp. 67–76, 2001. [25] Samijo, S.K., Willigers, J.M., Brands, P.J., Barkhuysen, R., Reneman, R.S., Kitslaar, P.J.,Hoeks,A.P., Reproducibility of shear rate and shear stress assessment by means of ultrasound in the common carotid artery of young human males and females, Ultrasound Med Biol. vol. 23, pp. 583–90, 1997. [26] Masaryk, A.M., Frayne, R., Unal, O., Krupinski,E., Strother,C.M., In vitro and in vivo comparison of three MR measurement methods for calculating vascular shear stress in the internal carotid artery, AJNR Am J Neuroradiol., vol. 20, pp. 237–45, 1999. [27] Tsuji, T., Suzuki, J., Shimamoto, R., Yamazaki, T., Nakajima, T., Nagai, R., Komatsu, S., Ohtomo, K., Toyo-Oka, T., Omata, M., Vector analysis of the wall shear rate at the human aortoiliac bifurcation using cine MR velocity mapping. AJR Am J Roentgenol., 178(4), 995–999, 2002. [28] Thomas, J.B., Milner, J.S., Rutt, B.K., Steinman, D.A., Reproducibility of image-based computational fluid dynamics models of the human carotid bifurcation. Ann Biomed Eng., 31(2), 132–41, 2003. [29] Steinman, D.A., Thomas, J.B., Ladak, H.M., Milner, J.S., Rutt, B.K., Spence, J.D., Reconstruction of carotid bifurcation hemodynamics and wall thickness using computational fluid dynamics and MRI. Magn Reson Med., 47(1), 149–59, 2002. [30] Taylor, C.A., Cheng, C.P., Espinosa, L.A., Tang, B.T., Parker, D., and Herfkens, R.J., In vivo quantification of blood flow and wall shear stress in the human abdominal aorta during lower limb excercise, Ann. Biomed. Eng., 30(3), 402–408, 2002. [31] Long, Q., Xu, X.Y., Bourne, M., Griffith, T.M., Numerical study of blood flow in an anatomically realistic aorto-iliac bifurcation generated from MRI data. Magn Reson Med., 43(4), 565–76, 2000. [32] Augst, A.D., Barratt, D.C., Hughes, A.D., Thom, S.A., Xu, X.Y., Various issues relating to computational fluid dynamics simulations of carotid bifurcation flow based on models reconstructed from threedimensional ultrasound images. Proc Inst Mech Eng [H], 217(5), 393–403, 2003. [33] Foo, T.K., Ho, V.B., Choyke, P.L., Contrast-enhanced carotid MR angiography. Imaging principles and physics, Neuroimaging Clin N Am., vol. 9, pp. 263–84, 1999. [34] De Marco, J.K., Schonfeld, S., Keller, I., Bernstein, M.A., Contrast-enhanced carotid MR angiography with commercially available triggering mechanisms and elliptic centric phase encoding, AJR, vol. 176, pp. 221–7, 2001. [35] Foo, T.K., Ho, V.B., Choyke, P.L., Contrast-enhanced carotid MR angiography. Imaging principles and physics, Neuroimaging Clin N Am., vol. 9, pp. 263–84, 1999. [36] Schoenberg, S.O., Knopp, M.V., Prince, M.R., Londy, F., Knopp, M.A., Arterial-phase threedimensional gadolinium magnetic resonance angiography of the renal arteries. Strategies for timing and contrast media injection, Invest Radiol. vol. 33, pp. 506–14, 1998. [37] Foo, T.K., Bernstein, M.A., Aisen, A.M., Hernandez, R.J., Collick, B.D., Bernstein, T., Improved ejection fraction and flow velocity estimates with use of view sharing and uniform repetition time excitation with fast cardiac techniques, Radiology, vol. 195, pp. 471–178, 1995. [38] Debatin, J.F., Davis, C.P., Felblinger, J., McKinnon, G.C., Evaluation of ultrafast phase-contrast imaging


441

in the thoracic aorta, MAGMA, vol. 3, pp. 59–66, 1995. [39] Lee, V.S., Spritzer, C.E., Carroll, B.A., Pool, L.G., Bernstein, M.A., Heinle, S.K., MacFall, J.R., Flow quantification using fast cine phase-contrast MR imaging, conventional cine phase-contrast MR imaging, and doppler sonography, in vitro and in vivo validation, AJR, vol. 169, pp. 1125–1131, 1997. [40] Pelc, L.R., Pelc, N.J., Rayhill, S.C., et al., Arterial and venous blood flow, noninvasive quantitation with MR imaging, Radiology, vol. 185, pp. 809–812, 1992. [41] Hamilton, C.A., Moran, P.R., Santago, P.N., Rajala, S.A., Effects of intra-voxel velocity distributions on the accuracy of the phase-mapping method in phase-contrst MR angiography, J Magn Reson Imaging, vol. 4, pp. 752–755, 1994. [42] Polzin, J.A., Alley, M.R., Korosec, F.R., Grist, T.M., Wang, Y., Mistretta, C.A., A complex-difference phase-contrast technique for measurement of volume flow rates, JMRI, vol. 5, pp. 129–137, 1995. [43] Lorensen, W.E., and Cline, H.E., Marching Cubes: A High Resolution 3D Surface Construction Algorithm, Computer Graphics, 21(3), 163–169, 1987. [44] Mitsuzaki, K., Yamashita, Y., Onomichi, M., Tsuchigame, T., Takahashi, M., Delineation of simulated vascular stenosis with Gd-DTPA-enhanced 3D gradient echo MR angiography: an experimental study. J Comput Assist Tomogr., 24(1), 77–82, 2000. [45] McInerney, T., Terzopoulos, D., Topology adaptive deformable surfaces for medical image volume segmentation, IEEE Trans Med Imaging, vol. 18, pp. 840–50, 1999. [46] Niessen, W.J., ter Haar Romeny, B.M., Viergever, M.A., Geodesic deformable models for medical image analysis, IEEE Trans Med Imaging, vol. 17, pp. 634–41, 1998. [47] Yim, P.J., Cebral, J.J., Mullick, R., Marcos, H.B., Choyke, P.L., Vessel surface reconstruction with a tubular deformable model, IEEE Trans Med Imaging. vol. 20, pp. 1411–21, 2001. [48] Yim, P.J., Vasbinder, B.C., Ho, V.B., Choyke, P.L., Isosurfaces as Deformable Models for Magnetic Resonance Angiography, IEEE Transactions on Medical Imaging, July, 2003. [49] Yim, P.J., Vasbinder, B., Ho, V.B., Choyke, P.L., A Deformable isosurface and vascular applications, Proceedings of SPIE, vol. 4684, pp. 1390–1397, 2002. [50] Yim, P.J., Cebral, J., Mullick, R., Choyke, P.J., Vessel Surface Reconstruction with a Tubular Deformable Model, IEEE Transactions on Medical Imaging, vol. 20, pp. 1411–1421, 2001. [51] Yim, P.J., Vasbinder, G.B.C., Ho, V.B., Choyke, P.L., A Deformable Isosurface and Vascular Applications, Proceedings of SPIE, Vol. 4684, pp. 1390–1397, 2002. [52] Cebral, J.R., and Löhner R., From Medical Images to Anatomically Accurate Finite Element Grids, Int. J. Num. Methdos Eng., 51, 985–1008, 2001. [53] Cebral, J.R., Löhner, R., Choyke, P.L., Yim, P.J., Merging of intersecting triangulations for finite element modeling, J Biomech., vol. 34, pp. 815–9, 2001. [54] Löhner, R., and Baum, J.D., Adaptive H-Refinement on 3-D Unstructured Grids for Transient Problems. Int. J. Num. Meth. Fluids, 14, 1407–1419, 1992. [55] Taubin, G., A Signal Processing Approach to Fair Surface Design, Computer Graphics Proceedings, 351–358, 1995. [56] Cebral, J.R., Yim, P.J., Löhner, R., Soto, O., Choyke, P.L., Blood Flow Modeling in Carotid Arteries Using Computational Fluid Dynamics and Magnetic Resonance Imaging, Academic Radiology, 9: 1286–1299, 2002. [57] Hoppe, H., Efficient Implementation of Progressive Meshes. Computers & Graphics, 22(1), 27–36, 1998. [58] Löhner R., Automatic unstructured grid generators. Finite Elements Analysis Design, vol. 25, pp. 111– 134, 1997. [59] Owen, S.J., A Survey of Unstructured Mesh Generation Technology. Proc. 7th International Meshing Roundtable, Sandia National Lab., 239–267, 1998. [60] Löhner, R., Cebral J.R., Parallel Advancing Front Grid Generation. In: Proc. 8th International Meshing Roundtable. South Lake Tahoe, California, 1999. [61] Löhner, R., Regridding Surface Triangulations, J. Comp. Phys., Vol. 126, pp. 1–10, 1996. [62] Löhner, R., Extensions and improvements of the advancing front grid generation technique. Communications in Numerical Methods and Engineering, 5, 119–132, 1996. [63] Mazumdar, J., Biofluid Mechanics. Singapore: World Scientific, 1992. [64] Kundu, P.K., and Cohen, I.M., Fluid Mechanics, 3rd edition, Elsevier, New York, 2004. [65] Wormersley, J.R., Method for the calculation of velocity, rate of flow, and viscous drag in arteries when the pressure gradient is known, J. Physiol., vol. 127, pp. 553–563, 1955.

442


[66] Taylor,C.A., Hughes, T.J.R., and Zarins, C.K., Finite Element Modeling of Blood Flow in Arteries. Comput. Methods Appl. Mech. Engrg., 158, 155–196, 1998. [67] Milner,J.S., Moore, J.A., Rutt, B.K., and Steinman, D.A., Hemodynamics of human carotid artery bifurcations: computational studies in models reconstructed from magnetic resonance imaging of normal subjects. J. Vasc. Surg. 28, 143–156, 1998. [68] Zhao, S.Z., Xu, X.Y., Hughes, A.D., Thom, S.A., Stanton, A.V., Ariff, B., and Long, Q., Blood Flow and Vessel Mechanics in a Physiologically Realistic Model of a Human Carotid Arterial Bifurcation. J. Biomech. 33, 975–984, 2000. [69] Cebral, J.R., Castro, M.A., Soto, O., Löhner, R., and Alperin, N., Blood-flow models of the circle of Willis from magnetic resonance data. Journal of Engineering Mathematics, 47, 369–386, 2003. [70] Cebral, J.R., Yim, P.J., Löhner, R., Soto, O., Marcos, H., and Choyke, P.L., New Methods for Computational Fluid Dynamics of Carotid Artery From Magnetic Resosnance Angiography, Proc. SPIE Medical Imaging, Vol. 4321, paper No. 22, San Diego, California, February 2001. [71] Quarteroni, A., Tuveri, M., and Veneziani, A., Computational vascular fluid dynamics: problems, models and methods. Report EPFL/DMA 11.98, 1998. [72] Thomaset, F., Implementation of finite element methods for Navier-Stokes equations. Springer-Verlag, 1981. [73] Gunsburger, M.D., and Nicolaides, R. (eds.), Incompressible Computational Fluid Dynamics: Trends and Advances. Cambridge University Press, 1993. [74] Hafez, M.M. (ed.), Numerical Simulation of Incompressible Flows. World Scientific, 2003. [75] Löhner, R., Applied CFD Techniques, John Wiley & Sons, New York, 2001. [76] Löhner, R., Yang, C., Cebral, J.R., Soto, O., Camelli, F., and Waltz, J., Improving the speed and accuracy of projection-type incompressible flow solvers. AIAA-03-3991-CP, 2003. [77] Löhner, R., Yang, C., Oñate, E., and Idelssohn, S., An unstructured grid-based, parallel free surface solver. Appl. Num. Math., 31, 271–293, 1999. [78] Saad, Y., Iterative methods for sparse linear systems. Boston: PWS Pub. Co., 1996. [79] Löhner, R., Yang, C., Cebral, J.R., Soto, O., Camelli, F., Baum, J.D, Luo, H., Mestreau, E. and Shadov, D., Advances in FEFLO, AIAA-02-1024, 2002. [80] Luo, H., Baum, J.D., and Löhner, R., A fast matrix-free implicit method for compressible flows on unstructured grids. J. Comp. Phys., 146, 664–690, 1998. [81] Sharov, D., Luo, H., Baum, J.D., and Löhner, R., Implementation of unstructured grid GMRES+LU-SGS method on shared-memory, cache-based parallel computers. AIAA-00-0927, 2000. [82] Soto, O., Löhner, R., Cebral, J.R., and Codina, R., A Time-Accurate Implicit Monolithic Finite Element Scheme for Incompressible Flow Problems, Proc. ECCOMAS CFD, Swansea, UK, September 2001. [83] Cebral, J.R., Putman, C., Pergolizzi, R., and Burgess, J.E., Patient-Specific Modeling of Atherosclerotic Carotid Arteries from Multi-Modality Image Data, Proc. IMECE’03, Washington DC, Nov. 16–24, 2003. [84] Estimation of Mechanical Stress on the Carotid Artery, Yim, P.J., Cebral, J., Löhner, R., Ho, V., Choyke, P.L., Annals of Biomedical Engineering, vol. 29, Supplement 1, Page S-77. [85] “Measurement of pressure drops at arterial stenoses from MR imaging” Abstract Book, Joint Meeting of the Biomedical Engineering Society and the IEEE Engineering in Medicine and Biology Society, Houston, Texas, (October 23–26, 2002), Yim, P.J. Cebral, J.R., Weaver, A., Lutz, R., Vasbinder, G.B.C., Choyke, P.L. p. 223. [86] Cebral, J.R., and Löhner, R., Flow Visualization On Unstructured Grids Using Geometrical Cuts, Vortex Detection and Shock Surfaces, AIAA-01-0915, 2001.


443

Numerical Modeling of Coronary Drug Eluting Stents Rosaire MONGRAIN a , Richard LEASK a , Jean BRUNETTE b , Iam FAIK a , Neil BULMAN-FELEMING a and T. NGUYEN a a McGill University, McGill, Canada b Université de Montréal, Montreal, Quebec, Canada Abstract. One of the principal therapies considered for the control of in-stent restenosis is the use of drug loaded polymer-coated stents for local delivery. We present two-dimensional and three-dimensional numerical models to study local delivery of drug eluting stents. The impact of various stent and flow parameters on the concentration distribution in the wall are investigated including the effect of the strut size, coating thickness, strut inter-distance and strut embedment in the vascular wall, blood flowing speed and the respective diffusion coefficients in the blood, wall and polymer. We also present criteria to assess the drug delivery efficiency based of the concept of the therapeutic window which aims at an spatial homogeneous concentration distribution and we introduce the variables to assess the amount of drug delivered in the wall. The results suggest that advection have a much stronger effect compared to diffusion in the blood media and that drug diffusivity in the arterial wall and in the polymer coating significantly affects the drug distribution. It is also shown that fully-embedded struts provide better spatial drug concentration uniformity after a short period of time and the half-embedded struts have a better temporal uniformity. Keywords. Drug eluting stent, numerical models, drug concentration, delivery efficiency

Introduction The success of coronary artery stenting has improved over the last decade. In stent restenosis has gone from approximately 40% at six months in the 1990s to around 10% at the turn of the century following the initial procedure1−4 . A number of promising methods for preventing restenosis have been investigated including systemic drug delivery5 , brachytherapy4 , local drug delivery6 , and targeted gene therapy7−9 . These therapies are meant to inhibit neointimal hyperplasia by altering the cellular growth response in the tissues surrounding the stent. In this chapter we will discuss some of the modeling tools developed to study drug delivery by drug eluting stents. In order to successfully prevent restenosis, drug eluting stents must deliver a therapeutic drug dose evenly through the treatment region for a defined duration. Favorable results have come from animal trials in porcine models6,10 , unfortunately information about the dose distribution over time is difficult to obtain from

444

R. Mongrain et al. / Numerical Modeling of Coronary Drug Eluting Stents

investigations of this nature. Moreover, porcine models of drug-eluting stents concentrate on measuring the degree of restenosis rather than drug concentrations in the tissue of interest. Clinical studies have shown impressive results with sirolimus-eluting stents even in complex disease vessels11−15 , however failure of the target vessel still occur in up to 10% in the first year. Numerical modeling of local delivery may help to identify some of the sources of these shortcomings and further optimize treatment. Current tissue transport models can be divided into two broad groups: porous medium models16,17 and diffusive models with various convective mechanisms incorporated into an effective diffusivity18,19 . These models use measured diffusivities and various porous model parameters in simulating transient drug delivery, often considering the multiphase structures of the arterial tissue. Results from the two basic models provide adequate agreement (ie. within the same order of magnitude) of drug delivery seen in vivo. Refinements to these models can be considered higher-order adjustments which will result in successive increases model accuracy, ultimately leading to more precise dose distribution in the arterial tissues. However tissue anisotropy and the complexity of shape and composition in actual pathological vessels could easily overshadow these higher-order model improvements. In this chapter, we present a relatively simple model for stent designers and cardiologists to study the influence of stent geometry, stent configurations, eluting polymer coating thickness and initial doses. With this in mind, we present a purely diffusive finite element model with appropriate boundary conditions, and focus our interest on representing some realistic geometric scenarios likely to arise following stent implantation. The numerical model is verified by comparing the impact of heparin-like (hydrophilic) and taxol-like (hydrophobic) diffusivities on drug concentration homogeneity distribution. The numerical tools developed can be used to: determine dose and dose-rate delivery for optimal therapy, compare dose delivery characteristics of polymer coated stent and quantify the effects of blood velocity, relative diffusion constants, coating thickness, inter-strut spacing, and strut apposition on dose delivery.

1. 2D Model of Drug Delivery The pharmacokinetic modeling of drug delivery from a polymer coated stent is a very complex problem. The complex three-dimensional geometry and structure of an atherosclerotic vessel with a stent are difficult to simulate. In addition, the polymer coating is relatively thin which makes numerical mesh generation difficult. The dispersion of the molecule is a multiphysics problem involving pulsatile fluid flow and mass transfer. In this section we present a very simple 2D steady flow model with rigid walls and a homogeneous wall structure to model drug delivery. In Section 3 we generalize the model in three-dimension and subsequently present an anatomically correct atherosclerotic vessel reconstructed from Intravascular Ultrasound (IVUS) image data. The proposed models (2D and 3D) are a simplification of the pharmacokinetics occurring in the context of arterial treatment. We assume homogeneous isotropic (nonporous) media. This means that the molecule transport due to transmural pressure gradient is neglected. The structures are assumed to be rigid and the flow steady. Similar models to investigate blood flow dynamics in stents have previously been proposed20,21 .


445

Figure 1. Description of the pharmacokinetic model around a single strut.

1.1. 2D Model Description The dispersion of the drug is modeled using the 2D advection-diffusion equation in the vessel lumen coupled to the Navier-Stokes equations and as a purely diffusive process in the arterial wall. It is assumed that the removal of drug at the outer radius (adventitial side) of the vessel is complete due to clearance by the vasa vasorum and lymphatic drainage, and therefore, modeled as a Dirichlet boundary condition. The model is illustrated in Fig. 1 and described with the following set of equations: ρ(u · ∇)u = −∇p + μ∇ 2 u ∇ ·u=0 ∂C + (u · ∇)C = DL ∇ 2 C ∂t ∂C = Dp ∇ 2 C ∂t ∂C = Dw ∇ 2 C ∂t

(1a) (1b) (1c) (1d) (1e)

where u is the blood velocity (m/s), C is the local drug concentration (Kg/m3 ), p the blood pressure (Pa), ρ the blood density (kg/m3 ), μ the blood dynamic viscosity (Pa.s) and DL ,Dp , Dw are the diffusion constants of the drug molecule in the lumen, the polymer and the arterial wall respectively (m2 /s). 1.2. Model Geometry The approach is illustrated with a simple helicoid stent geometry composed of a circular wire (similar to the CardioCoilT M stent from Medtronic) Fig. 2b. As a first approach, we consider a 2D representation of the symmetric plane along the axis of the stent (Fig. 2a). The dimensions adopted for this model are: • • • •

15 struts, 0.15 mm in diameter and located 0.7 mm center to center apart Polymer Coating Thickness = 0.005 mm Diameter of the artery = 3 mm Arterial wall thickness = 0.4 mm

446


Figure 2. A) Successive ring configuration corresponding to the 2D model B) Representation of the corresponding coil stent with the strut characteristics.

Figure 3. A) Typical mesh and B) a zoom of the mesh in between two struts (bottom), point 1 and 2 denote the location of where concentration levels were measured (see text).

1.3. Mesh Generation A finite element mesh was generated for the above configurations with quadrilateral linear elements using a combination of the pave and map schemes available in FIDAP (Fluent, USA). Smaller elements in the vicinity of the struts were used to better capture the physics in this region. This was accomplished by pre-meshing the sides and by creating boundary layers around the struts. A typical mesh is illustrated in Fig. 3. 1.4. Physical Parameters and Boundary Conditions From the governing equations, the principal physical model parameters are the mechanical properties of the blood, the properties of the coronary blood flow, the diffusivity of the drug molecule in the three continuum media and the initial concentration of the drug in the polymer coating. Blood properties used in this analysis are: density ρ = 1.057 g/cm3 and viscosity μ = 3.5 10−2 Poise. Physiological coronary blood flow rates vary between 0.5 and 1 cm3 /s, assuming fully developed flow condition at the inlet, this results in a maximum velocity (Vmax located at the center of the vessel) between 14 and 28 cm/s for an artery of 0.3 cm in diameter. The flow is considered laminar and fully developed hence the radial velocity distribution in the inlet is set to be parabolic:


447

Table 1. Summary of the boundary conditions. Boundary

Condition

Inlet

Parabolic velocity Profile and Concentration: C = 0

Endothelium

Velocity: Vx = 0 and Vy = 0

Advantitia

Concentration: C = 0

Axis of symmetry

Velocity: Vy = 0

V(R) = Vmax 1 −

R Ra

2 (2)

where R is the radial position and Ra the inner radius of the artery. The values of the diffusion constants of the drug in blood (DL ), in the polymer (Dp) and in the arterial wall (Dw) are unknown and depend mainly on the characteristics of the drug molecule and the type of the material properties of the polymer and vessel wall. Their range covers a wide spectrum, and hence their effect is studied for different orders of magnitude varying between hydrophobic and hydrophilic molecules. Molecule transport due to the transmural pressure gradient is neglected. This means that we consider the arterial wall and the polymer as non-porous media and thus all mass transport mechanisms are lumped into a single effective diffusion coefficient. A no-slip boundary condition is imposed at the solid walls (endothelium, stent struts). The concentration in the polymer is normalized so that the initial concentration is set to unity (Co = 1). This allows subsequent local concentrations in the polymer to be expressed directly as fractions of the initial concentration. Coronary arteries contain vasa vasorum and a lymphatic system which exchange biofluids within the vessel. This drainage establishes the boundary condition for the drug concentration at the outer limit of the arterial wall. Two extreme conditions are possible; 1) the molecule is completely washed away (C = 0 at the boundary) or 2) a saturation is achieved and the mass flux is zero (dC/dn = 0). As mentioned earlier, in these simulations it is assumed that there is complete removal at the outer limits of the artery wall (C = 0). The boundary conditions are summarized in Table 1. 1.5. Assessment of Drug Delivery Efficiency Two dimensional contour plots are a good way to visualize the distribution of drug concentration. In addition, numerical measures for comparing dose delivery between models were developed;

C dA

• The mean value of the concentration in the polymer ( dA ). • The local concentrations at point 1 and 2, Fig. 3. A is located at 250 microns from the endothelium directly facing the strut and B is in-between two struts. • The dose homogeneity index (DHI). • The remaining mass percentage (RMP). We define the DHI as the coefficient of variation, CV, of the concentration at each nodal point within the therapeutic region. A simple rectangular therapeutic region was considered with a depth of 200 μm into the arterial wall and a length of 17 mm (which extended beyond the first and last struts by 65 μm). The area taken up by the stent and

448


Figure 4. Illustration of 2D concentration distribution.

polymer are excluded from the calculation since these areas do not reflect on the success of local delivery. The mathematical form of the DHI is given as: M,N σC 1 Cij =√ −1 . DHI = CV ther.reg. = C C MN i,j

(3)

The lower the DHI the more homogeneous is the concentration distribution. The efficiency of drug transport to the vascular wall is measured with the Remaining Mass Percentage (RMP). The RMP is defined simply as the percentage of combined drug mass remaining in the stent and arterial wall after a specified time interval following stent implantation (the rest is lost in the blood stream). The value is given by:

Atotal i,j Cij dxi dyj = Cij (4) RMP = C0 Apolymer Apolymer i,j

When the dxi and dyj are made equal (by interpolation on a regular grid) there is no need for individual element weighting and the expression is simplified to the sum of all concentrations in the domain multiplied by the ratio of areas. The RMP provides a simple means of estimating initial concentrations necessary in achieving therapeutic dosages over a specified time period. It also offers a means of evaluating temporal dose distribution behavior, and its complement (1-RMP) indicates the fraction of initial mass removed from the system.

2. Effect of Model Variables on 2D Drug Delivery 2.1. Effect of Blood Velocity We first analyzed the effect of blood velocity on the drug diffusion rate and distribution patterns. The effect of blood velocity on the concentration patterns of the drug in the arterial wall was investigated at different steady flow velocities within the range of physiological coronary blood flow (with flow velocity varying between 10−3 to 28 cm/s). For these cases, the model assumes the stent struts to be half embedded with diffusion constants: Dw = 10−14 m2 /s, Dp = 10−13 m2 /s, DL = 10−9 m2 /s. An illustration of typical 2D concentration distribution distributions is shown in Fig. 4. The concentration results at different flow velocities are summarized in Table 2. Variations in blood velocity have a negligible effect on the relaxation time of the drug from the polymer as evaluated by the mean concentration in the polymer after one, two and three days. We also note that the local concentration in the arterial wall and hence the distribution pattern of the drug is not significantly affected by the variations in blood velocity for the entire range of diffusion coefficients.


449

Table 2. Effect of blood velocity on the mean normalized concentration in the polymer and in the vascular wall. Mean Normalized Concentration in Polymer

Mean Normalized Concentration in point 1 (See Fig. 3)

Maximum Velocity (mm/s)

1day

2days

3days

1day

2days

3days

280

8.30E-02

2.57E-02

1.20E-02

4.97E-02

3.86E-02

2.35E-02

200

8.30E-02

2.57E-02

1.20E-02

4.97E-02

3.86E-02

2.35E-02

140

8.30E-02

2.57E-02

1.20E-02

4.97E-02

3.86E-02

2.35E-02

10

8.30E-02

2.57E-02

1.20E-02

4.97E-02

3.86E-02

2.35E-02

1

8.30E-02

2.57E-01

1.20E-02

4.97E-02

3.86E-02

2.35E-02

1.00E-03

8.56E-02

2.67E-02

1.25E-02

5.00E-02

3.91E-02

2.39E-02

Figure 5. Final concentration in the polymer as a function of blood diffusion coefficient after 1 day, 2 days and 3 days.

2.2. Effect of the Drug Molecule Diffusion Coefficient in the Blood (DL ) The mean concentration of the drug molecule in the polymer after one, two and three days was evaluated for different values of the diffusion constant of the molecule in the blood (with DL varying between 10−12 to 10−6 m2 /s)which covers the range of diffusion constants between hydrophobic and hydrophilic molecules. The diffusion constants of the polymer and wall were fixed (Dw = Dp = 10−13 m2 /s). These values implicitly assumes that the drug molecule diffuses equally in all solid media. The results displayed in Fig. 5, show that the value of the mean concentration is unaffected by six orders of magnitude variation in the blood diffusion constant. 2.3. Effect of the Drug Molecule Diffusion Coefficient in the Polymer (Dp ) and in the arterial wall (Dw ) In the previous sections, we have shown little dependence of the concentration distribution on the blood flow velocity and on the diffusion coefficient in the blood. In this

450


Figure 6. Mean concentration remaining in the polymer after 1 day , 2 days and 3 days for different arterial wall diffusion coefficients.

section we investigate the effect of the relative values of the diffusion coefficients in the vascular wall (Dw ) and in the polymer coating (Dp ). To assess the effect of the different combinations of these two diffusion constants, we again evaluate both the mean concentration in the polymer and the ratio of the total concentration in the arterial wall to the C dA initial total concentration in the polymer: dA . The results are summarized in Table 3 and Figs 6a and 6b. Figure 6 presents the amount of the drug left in the polymer with time. Table 3 shows how much of this drug is accumulated in the arterial wall; the rest being washed away in the lumen and the outer radius of the model or still in the polymer. These results show a strong dependence of the drug release time and drug accumulation on both the drug diffusivity in the polymer and in the arterial wall. We can also note that better accumulation and homogeneity are achieved for lower values of the two diffusion constant (hydrophobic range). 2.4. Effect of Coating Thickness To investigate the effect of polymer thickness on the tissue drug concentration, the diffusion coefficients were fixed (Dp = Dw = 10−13 m2 /s) and blood advection was ignored (see Section 2.1). Thicknesses of 5, 10 and 15 μm were investigated for heparin-like diffusion coefficients. The time period used for these simulations was the 3 day duration of the previous sections. DHI results show similar homogeneities for the three poly-


451

Table 3. Effect of the relative values of Dw and Dp on the percentage of concentration in the arterial wall. Concentration (%) in Arterial Wall Dw

Dp

1day

2days

3days

1.00E-12

1.00E-12

3.66

0.27

0.02

1.00E-12

1.00E-13

5.43

5.87

5.42

1.00E-12

1.00E-14

5.24

18.76

7.09

1.00E-13

1.00E-12

11.04

5.73

3.02

1.00E-13

1.00E-13

26.28

18.95

12.32

1.00E-13

1.00E-14

23.51

22.09

17.32

1.00E-14

1.00E-12

4.94

3.28

2.62

1.00E-14

1.00E-13

18.95

18.66

17.20

1.00E-14

1.00E-14

21.91

27.19

29.51

Figure 7. Left: Concentration distribution (for two successive square struts) with a 5 μm coating (DHI = 0.65, RMP = 1.1%) Right: Concentration distribution (for two successive square struts) with a 15 μm coating (DHI = 0.73, RMP = 2.2%).

Figure 8. Top: Concentration distribution (shown around three successive struts for solid polymeric strut, Solid Stent: DHI = 1.344, RMP = 2.31% Bottom: Concentration distribution (coating thickness10 μm): DHI = 0.844, RMP = 0.73%.

mer layer thicknesses, with DHIs of 0.65, 0.66 and 0.73 for 5, 10 and 15 μm thickness. The remaining mass percentages (RMP) for the different layer thickness were; 1.1% for 5 μm, 1.5% for 10 μm and 2.2% for 15 μm. This suggests that in order to achieve the same drug retention in a 5 μm coating as a 15 μm coating, approximately twice the initial drug mass would need to be loaded within the polymer. Figure 7 illustrates the results for the 5 μm and the 15 μm thickness coatings. Although twice the drug remains in the system with the thickest coating, the distribution is less uniform.

452


Figure 9. Effect of inter strut spacing on drug concentration homogeneity (Top). Effect of inter strut spacing on remaining amount of drug in the wall (Bottom).

We also investigated the effect of having the entire stent made of drug eluting polymer and compared it with the 10 μm coating. The solid polymer stent has a slower elution rate and therefore a higher RMP but has a less even distribution of drug in the tissue. This suggests a coated approach with a higher initial dose of drug would provide good homogeneity of drug delivery and longer residence times. 2.5. Inter Strut Distance The impact of inter strut spacing on the concentration distribution was also investigated. Figure 9 summarizes the results. The homogeneity (DHI) of drug delivery increases with inter-strut spacing and the amount of drug remaining in the wall and coating (RMP) decreases as the inter-strut spacing increases due to changes in the spatial concentration gradient. As a result we conclude that the closer the struts, the more favorable the outcome. However there is a trade off, as increased stent contact area increases endothelial cell damage and the risk of inflammation and thrombosis. 2.6. Strut Embedment In the previous sections, we have shown that the physical variables that have the most influence on the concentration of distribution in the vascular wall are the relative values of the diffusion coefficients of the drug molecule in the wall and in the polymer (Dw, Dp). In this section, we investigate the impact of the strut embedment due to angioplasty for a given set of Dw, Dp. Different strut embedment configurations were compared on the basis of the uniformity of the drug concentration distribution over the arterial wall as well as the uniformity


453

Figure 10. Drug concentration distribution for the half embedded, non-embedded strut and fully embedded configurations. Table 4. Homogeneity indices and remaining mass percentages for four strut apposition scenarios. Apposition Scenario

DHI

RMP (%)

Completely embedded

0.74

2.2

Half embedded

0.66

1.5

Contacting

0.79

1.2

over time for Dw = Dp = 10−13 m2 /s. Again the objective in drug eluting-stent design and deployment is to achieve a uniform spatial and temporal distribution of the drug molecule. Using this model, three strut apposition configurations were investigated; no strut embedment, half-embedded struts and fully embedded struts. Figure 10 displays the general behaviour of the drug concentration distribution for all three configurations. The RPM is strongly affected by the degree of embedment. Interestingly, there seems to be an opposite DHI results for coated/solid (although coated is almost constant). Overall, the best concentration distribution is found with the fully embedded struts. However, again a balance between embedment and tissue injury should also be taken into account. The homogeneity parameters DHI and remaining mass percentage RMP as summarized in the following table.

3. 3D Dose Concentration Computations In this section, we present recent developments for modeling the pharmacokinetics of drug eluting stents in 3D. The complex 3D geometry requires the use of a Computer Assisted Design (CAD) program and a dedicated numerical mesh generation software to properly represent the artery and stent. The physics (blood flow and mass transport) is modeled with the same set of equations as in the 2D approach. The same concentration distribution criteria (DHI, RMP) are used to assess the drug dispersion. In addition, the surface contact area is calculated as a measure of endothelial disruption and resulting potential adverse biological response.

454


Figure 11. The Symbiotech stent and solid model.

Figure 12. 3D Stent strut geometry and numerical mesh.

Figure 13. 3D Wall shear stress distribution.

3.1. 3D Basic Geometry Numerical Model The 3D stent geometry of the Symbiotech stent (Laval, Qc) was generated using the CAD software Pro-Engineer (PTC, USA). Figure 12 shows the actual stent and its 3D CAD representation. The stent is shown in a 3.1 mm expanded configuration. The total stent length is 1.5 cm, the strut width is 63.5 μm and the modeled polymer coating thickness is 5 μm. 3.2. 3D Domain Discretization and Numerical Model For the 3D model, the discretization was carried out using ICEM-CFD (ANSYS, USA) with tetrahedral elements. Grid independence tests indicated that mesh element side lengths of 2.5 μm surrounding the polymer layer to 25 μm in the outer regions resulted in a favourable computation time/result agreement compromise. Figure 13 shows a magnified view near the strut of the solid model and local mesh used for simulation.


455

Coupled 3D Navier-Stokes and Advection-Diffusion equations were applied in the vessel lumen. The polymer coating and arterial wall were considered purely diffusive regions. Blood is considered an incompressible Newtonian fluid of density ρ = 1.05 g/cm3 and dynamic viscosity μ = 0.035 Poise (g/cm.s). ρ( u · ∇) u = −∇p + μ∇ 2 u

(5a)

∇·u=0

(5b)

Ct + ( u · ∇)C = Dlumen ∇ C 2

Ct = Dpolymer ∇ 2 C

(5c) Ct = Dartery ∇ 2 C.

(5d)

For comparison purposes, the same diffusivities used in 2D for hydrophilic and hydrophobic drug molecules, arterial wall, polymer coating and pathological tissues were implemented. Concentrations at the outer limit of the arterial wall and lumen inlet are set to zero for all time. At the outer wall limit this condition is meant to approximate the effect of dilution resulting from transport of mass from the vasa vasorum and lymphatic drainage. At the inlet concentrations due to dilution in blood will effectively be zero. No-slip velocity conditions are imposed on the polymer coating and arterial wall. Drug loading of the stent is represented as an initial concentration of 1.0 in the polymer. 3.3. 3D Velocity and Shear Stress Computation As mentioned in the introduction, non-physiological values of wall shear stress (WSS) are considered by some groups as a potential factor in the development of in-stent restenosis 20,21 . The 3D characterization of WSS and of flow patterns in a stented artery is therefore necessary to understand the phenomenon and to better design the next generation of stents. Indeed, 3D simulations of the stent allow us to investigate the global WSS distribution that cannot be obtained by simple 2D analyses. This is especially true for stents with complex shapes and for the description of the circumferential intra-struts WSS distribution. In the following figure we show recent 3D computations of the shear stress obtained from the velocity. We notice relatively high shear stress values at the struts (24.5 Pa) adjacent to a relatively low shear level in between the struts (1.4 Pa). This proximity of alternating shear stress levels could be involved in the pathogenesis of in-stent restenosis. In the following figure we show a preliminary 3D concentration distribution in the vascular wall and the corresponding remaining mass percentage (RMP) calculated over a period of one week. We propose an additional design criterion to take into account the biological response of the stent which we simply measure with the contact area of the stent with the vascular wall (this implicitly assumes a proportionality between the area of contact of the foreign body with the wall and the biological response). The area of contact index (ACI) is thus simply defined as the area of contact of the stent over the area of the cylinder defined by the stent length and circumference (this means that the ACI varies between 0 and 1). Finally, we combine all the design criteria into a single design efficiency criterion DE with the following weighted average:

456


Figure 14. 3D drug concentration distribution computation.

Figure 15. Left column: Modeled stent, middle column: corresponding numerical mesh, right column: summary of the dose distribution design efficiency criteria results.

7 DE = α

DH I −1 −1 DH Imin

8

+β

with α + β + γ = 1

F MP F MPmax

7

+γ

ACI −1

8

−1 ACImin

(6)

This means that DE varies between 0 (worst design) to 1 (best theoretical design) in terms of drug concentration homogeneity (DHI), remaining mass in the vascular wall (RMP) and biological response through contact area (ACI). We present in the following figure the analyses that were done for three different stent designs (Symbiotech, Cardiocoil and Palmaz). In these computations the values for αsgβ and γ were set to 0.5, 0.2 and 0.3 respectively to give more importance to the homogeneity equal to the combined effects of the remaining mass and area of contact. We note that in terms of the remaining mass percentage, the Palmaz stent appears to be the best. In terms of area of contact, the coil stent seems to perform best. Ongoing work includes the extension of the models to simulate the diffusion of the molecule into pathologic vessel structures taking into account the effect of the vascular matrix, lipid, fibrotic and calcific pools. These structures are reconstructed from Intravascular Ultrasound (IVUS) using manual segmentation as illustrated in Fig. 16.


457

Figure 16. Left to right: IVUS manual segmentation for the lumen and internal plaque structures; 3D CAD reconstruction of the lumen (gray), external elastic membrane (wireframe) and inclusions and Sent inserted in the 3D reconstructed pathologic wall structure.

The corresponding concentration computations are underway and should reveal how the drug transports through and in between the different plaque internal structures (matrix, lipid pools, fibrotic pools, calcific pools).

Discussion and Conclusions Although the success of coronary stenting has improved over the last decade, in-stent restenosis is still a common cause of stent failure. Numerical modeling of drug eluting stents present several advantages and have the potential to optimize stent design before testing in animals, therefore limiting costly experimental studies. In this chapter we presented some numerical tools to help assess and design drug eluting stents. We first showed that, for a simplified 2D model, changes in the blood velocity in the range of coronary blood flow do not significantly influence the rate of diffusion of the drug molecule from the polymer or its local concentration in the arterial wall. This suggests that the transport forces in the blood are mainly governed by advection while diffusion effects are negligible (Peclet number of the order 105 ). For the same reason, the results show little influence of the variability of the drug molecule diffusivity in the blood over several orders of magnitude on the total rate of diffusion from the polymer. The constants of diffusion of the drug molecule in the polymer and in the arterial wall were then shown to be the major parameters affecting the physiological transport of the drug after stent deployment. The evaluation of the drug distribution, more specifically its global homogeneity (spatial and temporal) over the arterial wall in a stent based delivery system is an important step to ensure that the drug will have the desired therapeutic effect. Previous studies have investigated the effect of different stent design parameters on specific aspects of the spatial distribution and have shown that circumferential and longitudinal strut spacing, significantly affects drug homogeneity18,19 . However, to the best of our knowledge, there are no numerical studies in literature concerning strut embedment. We have also shown that the essence of the 3D physics is recovered using 2D numerical models through the use of the introduced design efficiency drug delivery parameters, DHI, RMP and ACI allowing us to investigate the stent design in 2D before involved 3D computations. A method was presented to analyse the 3D distribution of a molecule for local delivery using realistic stent and vessel wall structures. The method allows visualization of the homogeneity of the 3D distribution of the molecule achieved in the vascular wall. It also provides a quantitative comparison of delivery efficiency in time

458


after stent implantation for various molecules and stent geometries. Such a tool is complementary to current animal investigations in assessing the efficiency of a given coated stent configuration.

References [1] Carrozza, J. P., Jr., Kuntz, R. K., Schatz, R. A., Leon, M., Goldberg, S., Savage, M. et al., Inter-series differences in the restenosis rate of Palmaz-Schatz coronary stent placement: differences in demographics and post-procedure lumen diameter. Cathet. Cardiovasc. Diagn., Vol. 31, pp. 173–8, 1994. [2] Lau, K. W., Ding, Z. P., Johan, A., Kwok, V., Lim, Y. L., Angiographic restenosis rate in patients with chronic total occlusions and subtotal stenoses after initially successful intracoronary stent placement. Am. J. Cardiol., vol. 83, pp. 963–10, 1999. [3] Schiele, F., Meneveau, N., Vuillemenot, A., Zhang, D. D., Gupta, S., Mercier, M. et al., Impact of intravascular ultrasound guidance in stent deployment on 6-month restenosis rate: a multicenter, randomized study comparing two strategies–with and without intravascular ultrasound guidance. RESIST Study Group. REStenosis after Ivus guided STenting. J. Am. Coll. Cardiol., vol. 32, pp. 320–8, 1998. [4] Schiele, T. M., Pollinger, B., Kantlehner, R., Rieber, J., Konig, A., Seelig, V. et al., Evolution of angiographic restenosis rate and late lumen loss after intracoronary beta radiation for in-stent restenotic lesions. Am. J. Cardiol., vol. 93, pp. 836–42, 2004. [5] Kubota, Y., Kichikawa, K., Uchida, H., Maeda, M., Nishimine, K., Makutani, S. et al., Pharmacologic treatment of intimal hyperplasia after metallic stent placement in the peripheral arteries. An experimental study. Invest Radiol., vol. 30, pp. 532–7, 1995. [6] Suzuki, T., Kopia, G., Hayashi, S., Bailey, L. R., Llanos, G., Wilensky, R. et al., Stent-based delivery of sirolimus reduces neointimal formation in a porcine coronary model. Circulation, vol. 104, pp. 1188–93, 2001. [7] Akyurek, L. M., Nallamshetty, S., Aoki, K., San, H., Yang, Z. Y., Nabel, G. J. et al., Coexpression of guanylate kinase with thymidine kinase enhances prodrug cell killing in vitro and suppresses vascular smooth muscle cell proliferation in vivo. Mol. Ther., vol. 3, pp. 779–86, 2001. [8] Hedman, M., Hartikainen, J., Syvanne, M., Stjernvall, J., Hedman, A., Kivela, A. et al., Safety and feasibility of catheter-based local intracoronary vascular endothelial growth factor gene transfer in the prevention of postangioplasty and in-stent restenosis and in the treatment of chronic myocardial ischemia: phase II results of the Kuopio Angiogenesis Trial (KAT). Circulation, vol. 107, pp. 2677–83, 2003. [9] Kipshidze, N., Moses, J., Shankar, L. R., Leon, M., Perspectives on antisense therapy for the prevention of restenosis. Curr. Opin. Mol. Ther., vol. 3, pp. 265–77, 2001. [10] Farb, A., Heller, P. F., Shroff, S., Cheng, L., Kolodgie, F. D., Carter, A. J. et al., Pathological analysis of local delivery of paclitaxel via a polymer-coated stent. Circulation, vol. 104, pp. 473–9, 2001. [11] Fattori, R., Piva, T., Drug-eluting stents in vascular intervention. Lancet, vol. 361, pp. 247–9, 2003. [12] Holmes, D. R., Jr., Leon, M. B., Moses, J. W., Popma, J. J., Cutlip, D., Fitzgerald, P. J. et al., Analysis of 1-year clinical outcomes in the SIRIUS trial: a randomized trial of a sirolimus-eluting stent versus a standard stent in patients at high risk for coronary restenosis. Circulation, vol. 109, pp. 634–40, 2004. [13] Lemos, P. A., Serruys, P. W., van Domburg, R. T., Saia, F., Arampatzis, C. A., Hoye, A. et al., Unrestricted utilization of sirolimus-eluting stents compared with conventional bare stent implantation in the “real world”: the Rapamycin-Eluting Stent Evaluated At Rotterdam Cardiology Hospital (RESEARCH) registry. Circulation, vol. 109, pp. 190–5, 2004. [14] Morice, M. C., Serruys, P. W., Sousa, J. E., Fajadet, J., Ban, H. E., Perin, M. et al., A randomized comparison of a sirolimus-eluting stent with a standard stent for coronary revascularization. N. Engl. J. Med., vol. 346, pp. 1773–80, 2002. [15] Schampaert, E., Cohen, E. A., Schluter, M., Reeves, F., Traboulsi, M., Title, L. M. et al., The Canadian study of the sirolimus-eluting stent in the treatment of patients with long de novo lesions in small native coronary arteries (C-SIRIUS). J. Am. Coll. Cardiol., vol. 43, pp. 1110–5, 2004. [16] Huang, Z. J., Tarbell, J. M., Numerical simulation of mass transfer in porous media of blood vessel walls. Am. J. Physiol, vol. 273, pp. H464–H477, 1997. [17] Qiu, Y., Tarbell, J. M., Numerical simulation of oxygen mass transfer in a compliant curved tube model of a coronary artery. Ann. Biomed. Eng, vol. 28, pp. 26–38, 2000.


459

[18] Hwang, C. W., Wu, D., Edelman, E. R., Physiological transport forces govern drug distribution for stentbased delivery. Circulation, vol. 104, pp. 600–5, 2001. [19] Lovich, M. A., Philbrook, M., Sawyer, S., Weselcouch, E., Edelman, E. R., Arterial heparin deposition: role of diffusion, convection, and extravascular space. Am. J. Physiol, vol. 275, pp. H2236–H2242, 1998. [20] LaDisa, J. F., Jr., Guler, I., Olson, L. E., Hettrick, D. A., Kersten, J. R., Warltier, D. C. et al., Threedimensional computational fluid dynamics modeling of alterations in coronary wall shear stress produced by stent implantation. Ann. Biomed. Eng, vol. 31, pp. 972–80, 2003. [21] Wentzel, J. J., Whelan, D. M., van der Giessen, W. J., van Beusekom, H. M., Andhyiswara, I., Serruys, P. W. et al., Coronary stent implantation changes 3-D vessel geometry and 3-D shear stress distribution. J. Biomech., vol. 33, pp. 1287–95, 2000.


461

Author Index Amores, J. Baldewsing, R.A. Bertrand, O.F. Bischof, H. Brunette, J. Budoff, M.J. Bulman-Feleming, N. Cai, J. Castro, M.A. Cebral, J. Chakareksi, J. Chen, Y. Christodoulou, C.I. de Korte, C.L. DeMarco, K. Duerk, J.L. Elbischger, P.J. Faik, I. Fei, B. Fox, M.D. Galaz, R. Griffin, M. Hatsukami, T.S. Holzapfel, G.A. Joshi, A. Kakkos, S. Kerwin, W. Klingensmith, J.D. Kyriacou, E. Lara-Montalvo, R. Laxminarayan, S. Leask, R. Lewin, J.S.

26 75 130 97 130, 443 148 443 55 412 412 208 384 241 75 412 384 97 443 394 1 130 241 55 97 130 241 360 300 241 208 1 130, 443 384

Li, C. Macione, J. Mastik, F. Mongrain, R. Nair, A. Nguyen, T. Nicolaides, A. Pattichis, C.S. Pattichis, M.S. Pedersen, P. Pujol, O. Radeva, P. Ranga, A. Regitnig, P. Rumberger, J. Salvado, O. Schaar, J.A. Serruys, P.W. Shinbane, J.S. Sonka, M. Suri, J.S. Tardif, J.-C. van der Steen, A.F.W. Vince, D.G. Wacker, F.K. Wahle, A. Wilson, D.L. Wu, D. Yang, Z. Yim, P. Yuan, C. Zhang, S.

1 1 75 130, 443 300 443 241 1, 241 241 208 276 26, 276 130 97 182 384 75 75 148 321 1, 384, 394 130 75 300 384 321 384, 394 1 1 412 55 384




Plaque Imaging: Pixel to Molecular Level (Studies in Health Technology and Informatics, Vol. 113) (Studies in Health Technology and Informatics)

Medical And Care 3 (Studies in Health Technology and Informatics) (Studies in Health Technology and Informatics)

Global Health Informatics Education (Studies in Health Technology and Informatics)

Medical and Care Compunetics 2 (Studies in Health Technology and Informatics, Vol. 114) (Studies in Health Technology and Informatics)

Research into Spinal Deformities 5 (Studies in Health Technology and Informatics) (Studies in Health Technology and Informatics)

Advances in Information Technology and Communication in Health - Volume 143 Studies in Health Technology and Informatics

International Perspectives in Health Informatics - Volume 164 Studies in Health Technology and Informatics

Connecting Medical Informatics and Bio-Informatics (Studies in Health Technology and Informatics)

E-Health: Current Situation And Examples Of Implemented & Beneficial E-Health Applications (Studies in Health Technology and Informatics) (Studies in Health Technology and Informatics)

From Grid to Healthgrid (Studies in Health Technology and Informatics, Vol. 112)

Regional Health Economies and ICT Services (Studies in Health Technology and Informatics Shti)

Evaluation Methods in Biomedical Informatics (Health Informatics)

The Conservative Scoliosis Treatment:1st SOSORT Instructional Course Lectures Book (Studies in Health Technology and Informatics) (Studies in Health Technology and Informatics)

eHealth Beyond the Horizon - Get IT There:Proceedings of MIE2008 (Studies in Health Technology and Informatics) (Studies in Health Technology and Informatics)

MEDINFO 2007: Proceedings of the 12th World Congress on Health (Medical) Informatics (Studies in Health Technology and Informatics)

Information Technology Solutions for Healthcare (Health Informatics)

Basic Engineering for Medics and Biologists (Studies in Health Technology and Informatics)

Information Technology in Health Care 2007 - Proceedings of the 3rd International Conference on Information Technology in Health Care: Socio-technical ... Studies in Health Technology and Informatics

Information Technology Essentials for Behavioral Health Clinicians (Health Informatics)

Challenges And Opportunities of Healthgrids: Proceedings of Healthgrid 2006 (Studies in Health Technology and Informatics)

Medical and Care Compunetics 4 (Studies in Health Technology and Informatics)

Research into Spinal Deformities (Studies in Health Technology and Informatics, 91)

Analysis, Design and Implementation of Secure and Interoperable Distributed Health Information Systems (Studies in Health Technology and Informatics, 89)

Transformation of Healthcare with Information Technologies (Studies in Health Technology and Informatics)

Biological Petri Nets - Volume 162 Studies in Health Technology and Informatics

CRC Studies in Informatics Series)

Telenursing (Health Informatics)

Medicine Meets Virtual Reality 18: NextMed - Volume 163 Studies in Health Technology and Informatics

Wearable eHealth Systems For Personalised Health Management: State Of The Art and Future Challenges (Studies in Health Technology and Informatics)

Biomedical Informatics: Computer Applications in Health Care and Biomedicine (Health Informatics)

Pediatric Informatics Computer Applications in Child Health

Plaque Imaging: Pixel to Molecular Level (Studies in Health Technology and Informatics, Vol. 113) (Studies in Health Technology and Informatics)

Medical And Care 3 (Studies in Health Technology and Informatics) (Studies in Health Technology and Informatics)

Global Health Informatics Education (Studies in Health Technology and Informatics)

Medical and Care Compunetics 2 (Studies in Health Technology and Informatics, Vol. 114) (Studies in Health Technology and Informatics)

Research into Spinal Deformities 5 (Studies in Health Technology and Informatics) (Studies in Health Technology and Informatics)

Advances in Information Technology and Communication in Health - Volume 143 Studies in Health Technology and Informatics

International Perspectives in Health Informatics - Volume 164 Studies in Health Technology and Informatics

Connecting Medical Informatics and Bio-Informatics (Studies in Health Technology and Informatics)

E-Health: Current Situation And Examples Of Implemented & Beneficial E-Health Applications (Studies in Health Technology and Informatics) (Studies in Health Technology and Informatics)

From Grid to Healthgrid (Studies in Health Technology and Informatics, Vol. 112)

Regional Health Economies and ICT Services (Studies in Health Technology and Informatics Shti)

Evaluation Methods in Biomedical Informatics (Health Informatics)

The Conservative Scoliosis Treatment:1st SOSORT Instructional Course Lectures Book (Studies in Health Technology and Informatics) (Studies in Health Technology and Informatics)

eHealth Beyond the Horizon - Get IT There:Proceedings of MIE2008 (Studies in Health Technology and Informatics) (Studies in Health Technology and Informatics)

MEDINFO 2007: Proceedings of the 12th World Congress on Health (Medical) Informatics (Studies in Health Technology and Informatics)

Information Technology Solutions for Healthcare (Health Informatics)

Basic Engineering for Medics and Biologists (Studies in Health Technology and Informatics)

Information Technology in Health Care 2007 - Proceedings of the 3rd International Conference on Information Technology in Health Care: Socio-technical ... Studies in Health Technology and Informatics

Information Technology Essentials for Behavioral Health Clinicians (Health Informatics)

Challenges And Opportunities of Healthgrids: Proceedings of Healthgrid 2006 (Studies in Health Technology and Informatics)

Medical and Care Compunetics 4 (Studies in Health Technology and Informatics)

Research into Spinal Deformities (Studies in Health Technology and Informatics, 91)

Analysis, Design and Implementation of Secure and Interoperable Distributed Health Information Systems (Studies in Health Technology and Informatics, 89)

Transformation of Healthcare with Information Technologies (Studies in Health Technology and Informatics)

Biological Petri Nets - Volume 162 Studies in Health Technology and Informatics

CRC Studies in Informatics Series)

Telenursing (Health Informatics)

Medicine Meets Virtual Reality 18: NextMed - Volume 163 Studies in Health Technology and Informatics

Wearable eHealth Systems For Personalised Health Management: State Of The Art and Future Challenges (Studies in Health Technology and Informatics)

Biomedical Informatics: Computer Applications in Health Care and Biomedicine (Health Informatics)

Pediatric Informatics Computer Applications in Child Health

Recommend Documents