PancDS in Real‐World Practice: A Prospective Multicenter Validation of a Clinical Decision‐Support System Bridging Experience Gaps in Pancreatic Lesion Diagnosis

Distinguishing pancreatic ductal adenocarcinoma (PDAC) from mass-forming pancreatitis (MFP) is challenging due to imaging mimicry and reader-dependent variability. PancDS is developed as a biomimetic pancreatic decision-support system that integrates clinical predictors, a radiomics signature, and self-developed deep features (PANet). PancDS is enabled by TriFusionNet, an adaptive fusion strategy designed to emulate expert reasoning by dynamically reweighting modalities according to diagnostic relevance. In a retrospective multicenter cohort of 1006 consecutive patients with pathologically confirmed resectable PDAC or MFP (2014-2023), 634 patients from Tongji Hospital were used for training/internal testing, and 372 patients from four independent hospitals for external testing. PancDS achieves internal and external AUCs of 0.936 (95% CI: 0.864-0.993) and 0.881 (95% CI: 0.833-0.924), respectively. In a reader study, PancDS significantly improves diagnostic accuracy and sensitivity, with the largest gains in intermediate and junior radiologists (P < 0.001) and minimal case-level deterioration (1.5-3.9%), functioning as a diagnostic equalizer. In a prospective consecutive cohort (Jan-Oct 2025; n = 151), PancDS achieves an AUC of 0.869 (95% CI: 0.725-0.978) and 94.7% accuracy. This practice-tested, prospectively evaluated system provides a reliable tool for PDAC-MFP differentiation, potentially informing surgical decision-making and enhancing diagnostic equity across diverse clinical environments.