Arterial Atherosclerosis Research In Medical Image Processing

Atherosclerosis, a vascular disease, is a major cause of mortality in modern societies. There are many research areas in studies of atherosclerosis, among which atherogenesis and characterization of vulnerable plaques are of great importance. First, the mechanisms of atherogenesis have not been fully elucidated. Abnormal hemodynamics and elevated endothelial permeability to low-density lipoproteins are associated with the genesis and development of atherosclerosis. They may also be accompanied by the change in endothelial morphology and vascular geometry, which can be visualized by medical imaging techniques. Second, atherosclerotic plaques are classified as vulnerable and non-vulneralbe plaques, while the latter lead to more severe clinical events. It has been demonstrated that plaque rupture with subsequent thrombus formation is responsible for most acute cardiovascular events. The plaques can be imaged, and thus their vulnerability can be assessed, which is helpful in preventing acute cardiovascular events. Medical imaging modalities, such as intravascular ultrasound, magnetic resonance, and fluorescence microscopy, play an important role in the aforesaid two research areas. However, they are restricted to limited use due to unequal development between hardware and software in the medical equipments. The software, i.e., approaches of medical image processing including image segmentation and feature extraction, tends to become the bottleneck of applying medical imaging modalities into further studies of atherosclerosis.The work presented in this dissertation aims to overcome the challenges that the traditional approaches of medical image processing encountered, and provide useful tools of computerized image analysis to support further elucidation of the mechanisms of atherogenesis and characterization of the vulnerable plaques. The study covers the aforementioned two areas, i.e., mechanisms of atherogenesis and characterization of vulnerable plaques.In the study of mechanisms of atherogenesis, the work falls into two parts.a) Investigation of the relationship between endothelial morphology and permeability based on microscopic images. A novel method is proposed to segment the fluorescence microscopic images and then extract morphological features. The speckle reducing anisotropic diffusion is adopted into the marker-controlled watershed to overcome the problems of low signal-to-noise ratios and blurred or even broken cellular borders. The optimal parameter settings are obtained from the cell detection receiver operating characteristic. After border detection, two categories of morphological features are then extracted, including cell shape features and intercellular features. Finally, regressions reveal that two intercellular features are correlated to albumin permeability (R>0.39 and P<0.0005). This finding is helpful in exploring the mechanisms responsible for high permeability.b) Investigation of the relationship between carotid bifurcation geometry and hemodynamics based on meganetic resonance images. First, lumen borders are detected by image segmentation and 14 geometric features are calculated. The factor analysis is then applied to these mutually correlated features to identify four mutually independent geometric factors. Finally, in multiple regressions, two factors alone are capable of predicting presumably atherogenic hemodynamics (R2>0.5 and P<0.0001). Two factors both have physical meanings, one measuring expansion at the bifurcation and the other measuring the colinearity of the common and internal carotid artery axes at the bifurcation. These results suggest that the geometric factors are significant predictors, and thus potential surrogates, of presumably atherogenic hemodynamics, which currenctly have to be calculated through complicated analysis of computational fluid dynamics. Therefore, the factors may be used in noninvasive screening of atherosclerosis in the future. In the study of vulnerable plaque characterization, the work is based onintravascular ultrasound, and it is also two-fold.a) Automatic image segmentation for atherosclerotic plaques in intravascular ultrasound images. In order to overcome the difficulties of traditional methods, i.e., contour initialization and robustness, a novel scheme is proposed to automatically and accurately detect lumen borders and external elastic membrances based on the contourlet transform and the active contour model. With the contourlet transform, an image is decomposed into lowpass components and bandpass directional subbands. The template matching is adopted in lowpass components to yield initial contours of the lumen border and the external elastic membrance. The anisotropic diffusion is then utilized in bandpass directional subbands to suppress noise as well as preserve vascular boundaries, and the contour evolution in the boundary vector field is used to obtain final contours.b) Feature extraction and classification for plaques in intravascular ultrasound images. First, all conventional morphological features of the plaques are extracted with full automation. Second, two categories of new features including texture and elastic features are also automatically extracted. The texture features consist of first-order statistics and features from the gray-level coocurrence matrix, and the elastic features are extracted from the strain tensor estimated by the nonrigid image registration. Finally, three types of features are used to design classifiers including Fisher linear discrimination, support vector machines, and generalized relevance learning vector quantization. The classification performance on 124 plaques, consisting of 36 vulnerable and 88 nonvulnearble ones, are compared between three classifiers by using leave-one-out cross validation. The results on test sets demonstrate that support vector machines outperform the other two with the sensitivity, specificity, correct rate, and Youden's index of 91.7%,97.7%, 96.7%, and 89.4%, respectively.