Automated Heart Wall Motion Abnormality Detection From Ultrasound Images using Segmental Knowledge

Coronary Heart Disease can be diagnosed by measuring and scoring regional motion of the heart wall in ultrasound images of the left ventricle (LV) of the heart. Studies have shown that the quality of diagnosis has great interand intra-observer variation, even among world-class expert cardiologists. This variability in diagnosis quality is particularly critical for CHD, because early diagnosis is a key factor in improved prognosis. We describe a completely automated and robust technique that detects diseased hearts based on automatic detection and tracking of the endocardium and epicardium of the LV. The local wall regions and the entire heart are then classified as normal or abnormal based on the regional and global LV wall motion. In order to leverage structural information about the heart we applied Bayesian Networks to this problem, and learnt multiple potential structures off of the data using different algorithms. We checked the validity of these structures using anatomical knowledge of the heart and medical rules as described by doctors. We then used a novel feature selection technique based on mathematical programming to add imaging features to the basic learnt structures. The resultant classifiers thus depend only on a small subset of numerical features extracted from dual-contours tracked through time. We verify the robustness of our systems on echocardiograms collected in routine clinical practice at one hospital, both with the standard cross-validation analysis, and then on a held-out set of completely unseen echocardiography images.