Unified CT–MR Foundation Model for Automated Crohn Disease Phenotyping and Prediction of Surgery: Results from the Nationwide EPI-IIRN Study

Ben Shaya ¹ Amit Yosef ² Gili Kurtser ^3,4 Leora Gotesdyner ^3,4 Elena Vainberg ^5,6 Anna Volkova ^5,6 Talar Hagopian ^5,6 Dotan Yogev ^3,6 Gili Focht ^3,6 Amir Ben-Tov ^7,8 Galia Zacay ^8,9 Eran Matz ¹⁰ Shir Zekaria ¹¹ Haggai Bar-Yoseph ^12,13 Hadas Ben-Atya ² Naama Gavrielov ² Sarah Teitz ¹⁴ Iris Dotan ¹⁵ Ruth Cytter-Kuint ^5,6 Dan Turner ^3,6 Moti Freiman ²

¹Efi Arazi School of Computer Science, Reichman University
²Faculty of Biomedical Engineering, Technion Israel Institute of Technology
³Juliet Keidan Institute of Pediatric Gastroenterology, Hepatology and Nutrition, The Eisenberg R&D Authority, Shaare Zedek Medical Center
⁴Faculty of Medicine, The Hebrew University
⁵Radiology Department, Shaare Zedek Medical Center
⁶Faculty of Medicine, The Hebrew University of Jerusalem
⁷Maccabi Research and Innovation Center, Maccabi Healthcare Services
⁸Gray Faculty of Medical and Health Sciences, Tel Aviv University
⁹Meuhedet Health Services Center, Meuhedet Health Services
¹⁰Leumit Health Services, Leumit Health Services
¹¹Clalit Research Institute, Clalit Health Services
¹²Rambam Health Care Campus, Department of Gastroenterology
¹³Bruce Rappaport Faculty of Medicine, Technion Israel Institute of Technology
¹⁴Faculty of Electrical and Computer Engineering, Technion Israel Institute of Technology
¹⁵Division of Gastroenterology, Rabin Medical Center

Purpose
To develop and validate a unified CT–MR foundation-model framework for automated Crohn disease phenotyping and individualized prediction of surgery from routine enterography.

Materials and Methods
In this IRB-approved retrospective multicenter study, CT enterography (CTE) and MR enterography (MRE) examinations were linked to the nationwide Israeli epi-IIRN registry. Ten radiologic findings extracted from reports by a validated language model weakly supervised a modality-conditioned DINOv3 model. The detection cohort comprised 7,170 examinations from 6,282 patients. The prognostic cohort comprised 5,148 examinations from 4,437 patients, including 782 surgeries. Predicted imaging phenotypes and 13 prespecified clinical variables were evaluated with elastic-net Cox models and a nonlinear discrete-time foundation-model head. An enriched clinical comparator added 96 pre-imaging laboratory and medication-history features. Validation included held-out testing, fully nested five-fold validation, leave-one-center-out validation, calibration, and decision-curve analysis.

Results
On held-out testing, macro-AUC was 0.780 (95% CI, 0.751–0.810) for MRE and 0.744 (95% CI, 0.697–0.789) for CTE. Compared with modality-specific models, the unified model showed similar MRE performance (0.780 vs 0.793) and higher CTE performance (0.744 vs 0.673; difference, 0.071 [95% CI, 0.001–0.143]). The primary imaging–clinical Cox model achieved surgery C-indices of 0.791 in the primary analysis and 0.783 (95% CI, 0.766–0.799) under fully nested validation. With enriched clinical variables, adding imaging increased the fully nested C-index from 0.731 to 0.789. The nonlinear discrete-time head achieved 0.813 in the primary analysis and 0.800 under fully nested validation. Leave-one-center-out C-index was 0.754 ± 0.053. Five-year observed and predicted surgery incidences were 13.4% and 13.8%, respectively; decision analysis showed net benefit across thresholds of 5%–50%.

Conclusion
A single CT–MR foundation model automated interpretable Crohn disease phenotyping and improved individualized prediction of surgery beyond clinical data, supporting scalable risk stratification from routine enterography.