Coronary artery disease is a major cause of death worldwide . If a coronary artery becomes narrowed or occluded owing to the buildup of plaque (e.g., calcium, fat, and cholesterol), or the formation of a thrombus, the blood flow to the myocardium will be reduced. Restriction of oxygenated blood flow is called ischemia, and the narrowing of a vessel is called stenosis.
In current clinical practice, conventional coronary angiography (CCA) via cardiac catheterization is considered to be the gold standard imaging technique to diagnose coronary artery disease . Computed tomographic angiography (CTA) is a potential alternative to CCA . CTA is a noninvasive technique that allows assessment of the coronary lumen and the evaluation for the presence of coronary calcifications and other causes of stenosis .
Recently, non-contrast CT has been used for mass screening for lung cancer [148, 286]. CT images are also useful in the quantification of coronary calcification . Coronary artery calcium is graded by Agatston score, volume, mass score, or density [37, 62] for risk stratification for future cardiac morbidity and mortality .
Coronary Artery Anatomy The coronary arteries supply oxygenated blood to the myocardium. The right and left main coronary arteries exit the ascending aorta from ostia just above the right and left aortic valve leaflets . These two branches subdivide and traverse the epicardium. The American Heart Association (AHA) Writing Group on Myocardial Segmentation and Registration for Cardiac Imaging divides the coronary arteries into 15-16 segments as shown in Fig. 3.66 .
Detection of Coronary Arterial Stenoses with CTA This section describes algorithms for segmentation of coronary arteries and detection of coronary arterial stenoses with CTA. Research on coronary artery segmentation have implemented several methodological solutions: topological thinning , particle filtering , graph-based analysis , fuzzy connectedness , vessel tracking and active contours , minimal cost path computation [205, 226], mathematical morphology , hybrid strategy using multi-scale filtering and Bayesian probabilistic approach with level set model , multi-scale enhancement and dynamic balloon tracking , and two-stage shape regression .
After segmentation, assessment for stenoses is performed. The approaches include a 3D level set [7,8], skeletonization and geometric analysis of a branch , morphological filtering and interactive masking , and fuzzy distance transform .
Fig. 3.66 Segments of coronary arteries.  RCA right coronary artery, RV right ventricular branch, AM acute marginal branch, PLV posterolateral ventricular branch, PDA posterior descending artery, LCA left coronary artery, LM left main coronary artery, LAD left anterior descending artery, DIAG1 first diagonal branch, DIAG2 second diagonal branch, LCX left circumflex artery, OM obtuse marginal branches
The segmentation algorithm by Schaap et al. is a coarse-to-flne robust shape regression approach . First, the method is initialized with an approximate centerline . The cross-sectional planes of vessels are generated based on the centerline. Next, vessels are represented by a combination of local SSMs of the vessels’ appearance and shape. Then, the coarse shape of the vessel is estimated with linear multivariate regression. Finally, the vessel shape is refined with nonlinear regression by optimizing the landmark coordinates to their most likely position.
Detection of Coronary Calcification from Non-contrast-Enhanced CT Images
This section describes detection algorithms for coronary calcifications from noncontrast CT images which are used in mass screening for lung cancer. The approaches of previously published methods involve a neural network  and two-stage classification with feature selection .
For the detection of coronary calcifications from thick-section CT images (10 mm slice thickness), Ukai et al. proposed an algorithm that is composed of four processes . First, a CT image set is divided into three volume sections (upper, middle, and lower thirds of the heart) by a neural network (four inputs: slice position, heart shape, scapula, CT value uniformity), which is trained using the back propagation algorithm. Second, each section is segmented, using the information from the adjacent lung and vertebral body. Third, the candidate regions for the coronary calcifications are detected using a weight coefficient map consisting of a prior probability of the location for the coronary artery as shown in Fig. 3.67.
This probability is determined based on the distribution of the coronary arteries, which were manually segmented from 80 patients’ heart regions normalized in terms of height and width. Finally, the artifact regions included in the candidate regions are excluded by the diagnostic rule based on a neural network.
Fig. 3.67 Weight coefficient map. (a) Weight coefficient map in upper part of the heart, (b) weight coefficient map in middle part of the heart, (c) weight coefficient map in lower part of the heart. The value is a prior probability of the location of the coronary artery
For the detection of coronary calcifications from thin-section CT images (3 mm slice thickness), Isgum et al. proposed an algorithm that is composed of three processes . First, candidate objects are extracted using threshold of the intensity and size. Second, a set of features is calculated. This set is composed of volume, shape features (three eigenvectors, Xb X2, X3), spatial features (the coordinate system is defined by determining the smallest box around the heart), and appearance features (maximum intensity, average intensity, derivatives (Lx, Ly, Lz, Lxx, Lxy, Lxz, Lyy, Lyz, Lzz). Finally, candidate objects are classified into either positive or negative objects using a two-stage classification system with a k-nearest neighbor classifier and a feature selection scheme (sequential floating forward feature selection[SFFS]) . In this study, 14 efficient features (six spatial and eight appearance features) were employed by the feature selection.
This section describes technologies based on computational anatomy model for the diagnosis of coronary artery disease. Clinical applications including these algorithms can improve the diagnostic performance of coronary artery disease.