Abstract / Summary
Timely decision support in emergency imaging requires robust integration of structured clinical variables and contrast-enhanced CT while minimizing reader-dependent processing and improving reproducibility. We propose a modular segmentation-to-fusion framework for multimodal clinical–CT prediction using acute pancreatitis as a case study. The pipeline uses automated pancreas ROI generation with post-segmentation quality control, extracts complementary radiomics and frozen deep features, and integrates multimodal evidence through a calibrated stacked ensemble. We evaluated the framework on a retrospective single-center contrast-enhanced CT cohort using internal training and held-out test splits, assessing discrimination, calibration, and decision-curve performance. The proposed multimodal stacked ensemble achieved an AUC of 0.913 (95% CI 0.840–0.972) on the internal held-out test cohort and showed improved calibration and decision-curve performance compared with comparator models. These results suggest that reproducible multimodal clinical-imaging fusion may provide complementary information for internal risk estimation of AP in a single-center retrospective setting.