Learning vascular continuity for faithful liver vessel segmentation

Spatial continuity is the primary characteristic of liver vessels, which can be helpful in bridging gaps between vessel segments and locating very thin vessels in liver scans. As liver images are typically acquired in the transverse plane, vessels appear as a mixture of blob-like and tube-like structures, making their continuity difficult to preserve. We therefore employ recurrent modeling to infer vessel continuity from the three orthogonal anatomical views, enabling long-range directional information propagation within each view. The inferred continuity is used in two-fold: (1) shaping feature learning on vessel-like regions and (2) balancing the skewed distribution of vessels against others. Comprehensive experiments on three well-established datasets, namely 3D-IRCADb, MSD, and LiVS, show that the proposed model achieves relative improvements of up to 4.02% in Dice and 3.27% in centerline Dice over the corresponding second-best methods. Furthermore, our method improves the recovery and continuity preservation of ambiguous and fragmented vessels, providing a computationally efficient solution for liver vessel segmentation.