An Explainable Multi-Branch Decision-Support Framework for Early Pancreatic Cancer Detection Using Clinical Biomarkers and CT/MR Imaging Evidence

Abstract

Pancreatic ductal adenocarcinoma (PDAC) requires earlier diagnostic support, but medical artificial intelligence systems must remain leakage-aware, clinically interpretable, and transparent about validation boundaries. This study developed an explainable multi-branch decision-support framework combining a validated clinical biomarker model with CT imaging evidence, external healthy-control safety analysis, and cross-modal annotation support. The primary patient-level branch was a CatBoost model using age, sex, creatinine, plasma CA19-9, LYVE1, REG1B, TFF1, and REG1A, with leakage-prone variables excluded. CT Grand Jam was evaluated as a tumor-suspicious slice/slab imaging evidence branch. NIH Pancreas-CT was used for healthy-control domain-shift and safety analysis. MRI/PANTHER and CURVAS-PDACVI were used for tumor-pattern localization, segmentation-based explainability, and annotation-aware support, not diagnostic validation. Clinical CatBoost achieved ROC-AUC = 0.9513, PR-AUC = 0.9058, sensitivity = 0.9146, specificity = 0.8645, accuracy = 0.8814, and Brier score = 0.0812. CT Grand Jam achieved ROC-AUC = 0.8478 and PR-AUC = 0.7167. NIH safety filtering on 80 healthy controls produced 68/80 negative, 8/80 review, and 4/80 suspicious outputs. MRI/PANTHER included 142 readable image/label rows, median tumor volume = 13390.09 mm³, and OOD reconstruction error mean = 0.012183. The final layer implements deterministic decision-level integration into Low Risk, Review-Indeterminate, and High Suspicion categories. The framework is transparent, reporting-aware, and clinically interpretable, but prospective matched-cohort validation remains required before deployment.