Predicting Stenosis Severity and Localization in Coronary Artery using Deep Learning

Coronary artery disease (CAD), in particular, stenosis, has been one of major causes of death from cardiac disease. Detecting the severity of stenosis, as well as localizing the distribution of stenosis in coronary artery have been urgent technical problems to be resolved. In this paper, we focus on these two issues; severity of stenosis and localization of stenosis in coronary arteries. Given contrast CT images of stenosis patients as well as able-bodied, we have coped with these problems. For the severity of stenosis, we transform contrast CT images to MPR (Multi-Planar Reconstruction) images as pre-processing, then we formulate the problem as multi-label multi-class classification problem. For the localization of stenosis, we have asked doctors and radiologists to provide us with the segment numbers (e.g., RCA-1, RCA-2, LAD-7, LCX-13) as well as central lesion image within contrast CT images for each patient. We then formulate the problem, similar to severity prediction, as multi-label, multi-class classification problem. For the localization system, we produce the interval of CT slice numbers which is most likely to include the central lesion image. Through our experiments, we demonstrate that our proposed system is very effective and promising in both problems.