Detection, localization and quantification of non-calcified coronary plaques in contrast enhanced CT angiography

State-of-the-art imaging equipment has increased clinician's ability to make non-invasive diagnoses of coronary heart disease (CHD); however, high volumes of imaging data make manual abnormality detection cumbersome in practice. In addition, the interpretation of CTA heavily relies upon the previous knowledge of the clinician. These limitations have driven an intense research in the context of automated solutions for fast, reliable and accurate diagnosis. Accordingly, in this thesis, we present an automated framework for detection, localization and quantification of the non-calcified coronary plaques in cardiac computed tomography angiography (CTA). The first contribution of the thesis is a coronary segmentation algorithm that is adaptive to the contrast agent and employs a hybrid energy incorporating local and global image statistics in a segmentation framework using partial differential equations (PDEs). Accordingly, we illustrated with the help of experimental evidence that a volume-specific intensity threshold leads to an improved segmentation in CTA. In the subsequent step, we employed a hybrid region-based energy for improved segmentation in CTA imagery. The hybrid energy couples an intensity-based local term with an efficient discontinuity-based global model of the image for optimal segmentation. The proposed method is less sensitive to the local optima problem and helps in reducing false positives, as well as it allows a certain degree of freedom for the initialization. Moreover, we employed an auto-correction feature for improved segmentation, as an auto-corrected mask captures the emerging peripheries of the coronary tree during the curve evolution. The effectiveness of the proposed model is demonstrated with the help of both qualitative and quantitative results, with a mean accuracy of 80% across the CTA dataset. The capability to address the variations in initial mask and localization radii simultaneously, makes our algorithm a feasible choice for coronary segmentation. The second contribution of the thesis is an automatic approach to analyse the segmented coronary tree for the presence of non-calcified plaques. The specific focus of this work is detection of non-calcified plaques in CTA, as intensity overlap between blood, fat and non-calcified plaques make the detection challenging. Non-calcified plaques are identified based on mean radial profiles that average the image intensities in concentric rings around the vessel centreline. Subsequently, an SVM classifier is applied to differentiate the abnormal coronary segments from normal ones. A total of 32 CTA volumes have been analysed and a detection accuracy of 88.4% with respect to the manual expert has been achieved. For plaque-affected segments, we further proposed a derivative-based method to localize the position and length of the plaque inside the segment. The plaque localization accuracy has been around 83.2%. Moreover, the proposed model has been tested on three different CTA datasets and has produced consistent results, demonstrating its reproducibility for generic CTA data. The final contribution of the thesis is a method to segment and quantify the non-calcified plaque. After evaluating the vessel wall thickness, posterior probability based voxel classification has been performed to quantify the lumen and plaque, respectively. Both qualitative and quantitative results demonstrate that the proposed model shows a good agreement with three independent experts. To optimize the processing time, we employed sparse field method in a level-set based active contour evolution