VesselVision-Net: Hybrid Global-Local Context Learning for Liver Vessel Segmentation in Hepatic Surgery Preplanning

Accurate segmentation of liver vessels on contrast-enhanced CT is essential for safe hepatic surgery preplanning. It is still challenging because vessels vary greatly in size and have complex branching, and small branches are difficult to detect due to their low contrast. In this work, we propose VesselVision-Net, a hybrid global-local context learning framework for liver vessel segmentation, which focuses on detecting small vessel branches. The model uses a dilated fusion module (DFM) to capture better small vessel branches and a paired attention module (PAM) to enhance vessel-focused spatial and channel-wise attention. DFM aggregates multiscale features without losing resolution in small branches. PAM refines spatial and channel-wise information to emphasize vessel structures and suppress background. We design a composite loss function combining weighted Dice, power-weighted Dice, and cross-entropy to improve sensitivity to small and low-contrast vessels. While recent CNN- and transformer-based baselines can reliably segment the major vessel trunks, they still struggle to detect small peripheral branches, the primary challenge in liver vessel segmentation. In contrast, VesselVision-Net is specifically designed to enhance sensitivity to these small, low-contrast structures. VesselVision-Net is evaluated on the public 3D-IRCADb-01 and on a clinically annotated dataset. Across both datasets, the model achieves a Dice score of about 0.71 and a sensitivity of about 0.7313. Further analysis shows an ROI-based Dice score of 0.84 on small branches. Expert reviewers rated the clinical usability as "high'' in 77% of cases. These results suggest that VesselVision-Net can provide reliable vascular maps for hepatic surgery preplanning and computer-assisted decision support.