Deep Learning Coronary Artery Centerlines Mapping from Contrast-Enhanced CT Images of the Heart

Cardiovascular disease, and coronary heart disease in particular, is the number one killer in both Europe and the United States. Extraction of the coronary artery centerline tree from contrast-enhanced cardiac CT images is a challenging prerequisite for subsequent non-invasive functional assessment of the arterial tree. A coronary artery tree centerline tracker based on a novel multi-path deep convolutional neural network is presented. The network estimates the vessel radius and two opposing moving directions at each location, using the direction classification output to aid vessel radius regression. With only two starting points located anywhere in the left and right arterial tree respectively, the proposed tracker has the capability to extract the complete coronary artery tree. Automatic post-processing transforms the extracted cloud of points into a graph-based map in which the coronary ostia, arterial intersections, and endpoints are properly encoded. The MICCAI 2008 Coronary Artery Tracking Challenge training dataset was used for training and testing in a leave-one-patient-out cross-validation framework. Experimental results show that the proposed approach performs better than the reference state-of-the-art solution obtaining an average distance between reference centerline and the two placed direction points of 0.31 mm and an average accuracy inside score (AI) of 0.32 mm (15% and 49% performance increase, respectively), tracking the whole arterial tree consistently starting from just 2 seeds.