Predicting Stenosis Severity of Coronary Arteries with Non-Contrast CT and MPR Images Using Deep Neural Networks

In this paper, we present our investigations on methods toward predicting stenosis severity from two types of coronary artery images; Non-contrast CT images and MPR (Multi-Planar Reformatted) images converted from contrast CT images. For non-contrast CT images, we introduce 3D DNNs pretrained with Kinetics movies, while for MPR images, we first generate synthesized MPR images, then we apply 2D DNNs to attempt to estimate the stenosis severity of coronary arteries. In both approaches, we adopt QCA (Quantitative Coronary Angiography) values ranging from 0.0 (complete healthy) to 100.0 (complete stenosis, or occlusion) as the indicator of stenosis severity. We formulate the QCA prediction as a three-class classification problem (i.e., class 0 for [0.0,50.0), class 1 for [50.0, 75.0), and class 2 for [75.0,100.0]), as well as a regression problem. The regression is used only when the stenosis is positive (i.e., QCA ≥ 50.0) for both approaches. Through our experiments, we demonstrate that both approaches (i.e. 3D DNN and 2D DNN approaches) resulted in favorable performance in terms of AUC (Area Under Curve) in classification and RMSE (Root Mean Squared Error) in regression. We believe that our methods have potentials to help doctors and radiologists with real medical diagnosis, because of the reasonably good predictability of stenosis severity of coronary arteries.