A Hybrid Approach for Coronary Artery Anatomical Labeling in Cardiac CT Angiography

Abstract

Automated anatomical labeling plays an essential role in the effort of developing a computer-aided diagnosing system for coronary artery diseases. The large variation between individuals, presence of vascular stenosis or blockages, and possible image degradations makes the problem extremely challenging. In this paper, we propose a hybrid approach that combines the strength of traditional rule-based methods and recent deep learning methods, ensuring interpretability and consistency of the labeling process while being capable of learning parameters in a data-driven scheme. Our method is composed of two major components. First, we present our method of artery tree building from possibly noisy initial segmentation. The proposed algorithm links missing vessel segments while reducing noise, resulting in a complete and clean centerline based artery tree representation. Next, our labeling algorithm combining gated graph convolutional network (GGCN) and logical rules is elaborated. Experiments have demonstrated encouraging results both for completeness of artery tree building and accuracy of anatomical artery labeling.