Segmentation to Label: Automatic Coronary Artery Labeling from Mask Parcellation

Zhuowei Li, Qing Xia, Wenji Wang, Zhennan Yan, Ruohan Yin, Changjie Pan, Dimitris N. Metaxas

Abstract: Automatic and accurate coronary artery labeling technique from CCTA can greatly reduce clinician’s manual efforts and benefit large-scale data analysis. Current line of research falls into two general categories: knowledge-based methods and learning-based techniques. However, no matter in which fashion it is developed, the formation of problem finally attributes to tree-structured centerline classification and requires hand-crafted features. Here, instead we present a new concise, effective and flexible framework for automatic coronary artery labeling by modeling the task as coronary artery parsing task. An intact pipeline is proposed and two paralleled sub-modules are further designed to consume volumetric image and unordered point cloud correspondingly. Finally, a self-contained loss is proposed to supervise labeling process. At experiment section, we conduct comprehensive experiments on collected 526 CCTA scans and exhibit stable and promising results.