Joint segmentation of pancreatic tumors and peripancreatic vessels via a clinical risk-aware uncertainty alignment framework

Preoperative assessment of pancreatic cancer resectability critically relies on accurate joint segmentation of tumors and peripancreatic vessels from contrast-enhanced CT. However, beyond segmentation accuracy alone, effective decision support in practice further requires task-specific characterization of peri‑tumoral infiltration risk and topological discontinuity risk. To address this, we propose a clinical risk-aware uncertainty alignment framework that achieves reliable risk characterization while maintaining high segmentation accuracy. The framework is built upon a dual-decoder architecture with evidential deep learning for joint segmentation and uncertainty estimation. Two clinically grounded alignment strategies are further designed to transform this uncertainty into task-specific risk indicators. Specifically, an infiltration-aware boundary uncertainty alignment module is proposed, which synthesizes texture, spatial, and vascular proximity factors into an infiltration-inspired prior and uses its ordinal structure to guide alignment, yielding infiltration-risk-aware uncertainty from which infiltration-aware tumor boundaries are derived for downstream invasion analysis. A topology-aware discontinuity uncertainty alignment module is further designed, which uses skeleton-mismatch regions as a connectivity-related proxy and constrains uncertainty by a three-tier ordinal relation with region-weighted supervision, yielding discontinuity-risk-aware uncertainty and improved vessel connectivity. Multi-center validation on 1058 patients from five medical centers demonstrates favorable and stable cross-institutional performance, and downstream experiments explore the utility of the aligned uncertainty for tumor-vessel invasion analysis. Overall, this study couples accurate tumor and vessel delineation with clinically grounded risk characterization, with the potential to support tumor-vessel assessment in preoperative planning. Our source code will be released at https://github.com/SJTUBME-QianLab/CRUA-Net.