TaG-Net: Topology-Aware Graph Network for Vessel Labeling

Abstract

Anatomical labeling of head and neck vessels is a vital step for cerebrovascular disease diagnosis. However, it remains challenging to automatically and accurately label vessels in computed tomography angiography (CTA), since head and neck vessels are tortuous, branched, and often close to nearby tubular-like vasculatures. To address these challenges, we transform the voxel labeling problem into the centerline labeling task and propose a novel method of topology-aware graph network (TaG-Net) for vessel labeling of 18 segments covering both head and neck, in which the efficiency of centerline’ sparse representation using the point cloud is exploited and vessel’s topological structure can be better represented using the topology-aware graph. First, a topology-aware graph is constructed from the extracted vessel centerlines. Second, we design topology-preserving sampling and topology-aware feature grouping so that the network’s sampling and grouping layers preserve the vascular structures. Third, the vascular features extracted from the point processing layer and the GCN layer are aggregated for centerline labeling. Finally, the labeling task is accomplished by assigning the closet label from each point of the centerline to the mask voxels. Using head and neck CTA of 401 subjects and a five-fold-cross-validation strategy, experiments show that TaG-Net yields an average recall of 0.977 and an average precision of 0.977, with mean F1 as 0.977 for centerline labeling. After back-propagating labels onto vessel masks, TaG-Net achieves an average Dice coefficient of 0.991 for 18 vessel segments compared to that of 0.980 by the V-Net. The results indicate that the proposed network could facilitate head and neck vessel analysis by providing automatic and accurate vessel labeling.