Fig. 1: Overview of the PAN-VIQ framework for automated quantification of vascular invasion in pancreatic cancer and study design.

A Overall pipeline of PAN-VIQ: The model receives dual-phase (arterial and portal venous) CT images as input, performs tumor–vessel segmentation, and computes continuous encasement angles at tumor–vessel interfaces. Outputs include both continuous values (e.g., 38°) and categorical classifications (e.g., contact <180°). Model performance is benchmarked against intraoperative findings and radiologist interpretation. B Step 1: Training and validation of the three-stage segmentation model. The model was iteratively trained via a hybrid strategy comprising supervised pretraining, semi-supervised tuning, structural refinement, and post-processing. Manual review and correction by radiologists were performed after each stage to improve robustness and anatomical accuracy. C Step 2: Establishment of the vascular invasion quantification model. The vessel centerline (AB) was extracted. For each centerline point O, a local 2D plane CDEF orthogonal to the centerline tangent at O was generated, and the tumor/vessel masks were resampled onto this plane to obtain cross-sections. On CDEF, the boundary intersections between the tumor and vessel cross-sections were labeled b and c, and the centroid of the tumor cross-section was denoted a. The wrapping angle was defined as ∠bac, representing the degree of circumferential tumor involvement. D Step 3: Evaluation and validation of PAN-VIQ. Retrospective validation included 1759 internal cases and 164 external cases across three hospitals. Prospective validation was performed in 202 newly recruited patients, with model predictions compared to assessments by junior and senior radiologists, and benchmarked against surgical findings. E Patient cohort flowchart: The training, validation, and testing datasets were derived from 2130 patients with pathologically confirmed PDAC. Inclusion and exclusion criteria are detailed for both internal and external cohorts.