Joint Coronary Centerline Extraction And Lumen Segmentation From Ccta Using Cnntracker And Vascular Graph Convolutional Network

Automatic analysis of coronary artery in coronary computed tomography angiography (CCTA) is important for clinicians to diagnose and evaluate coronary artery disease (CAD). Often there are two analysis tasks involved: centerline extraction and lumen segmentation, which are related yet often treated separately in the literature. In this work, we leverage their mutual relationship and propose an automatic approach for joint centerline extraction and lumen segmentation from CCTA using a hybrid of deep learning models. Our approach features the following designs, including: (i) the use of CNNTracker that traces out the centerline from seeds to detected ostia or the end of vessels, which fixes the breakage issue commonly found in previous pixel-based segmentation methods; (ii) a vascular Graph Convolutional Network (GCN) that leverages the geometric shape prior for accurate mesh-based lumen segmentation; (iii) the alternation between CNNTracker and GCN that refines the centerline and reduces the drift in tracking. We experiment the proposed method on CCTA data from 101 patients and achieve state-of-the-art performance in both centerline extraction and lumen segmentation of the right coronary artery (RCA), left anterior descending (LAD), and left circumflex (LCX).