Towards personalized management of intraductal papillary mucinous neoplasms with multimodal artificial intelligence.

e16461 Background: According to the Fukuoka and Kyoto consensus guidelines, the clinical management of pancreatic intraductal papillary mucinous neoplasms (IPMNs) primarily depends on imaging features, cytology, and clinical variables such as CA19-9 serum levels. While these guidelines are sensitive to high-grade or invasive IPMN, they lack specificity, resulting in surgical overtreatment. We propose a clinically applicable multimodal deep learning model to accurately predict the optimal management approach, i.e., surgical resection or surveillance, based on imaging and clinical data. Methods: This retrospective study included 180 patients with IPMN who underwent surgical resection at Indiana University School of Medicine. We developed prediction models for the most optimal management - surgical resection (for high-grade and invasive IPMN) vs surveillance (for low-grade IPMN) - based on individual preoperative magnetic resonance imaging (MRI) sequences (T1 with and without contrast, T2) and clinical data (age, sex, race, BMI, CA19-9, cyst size, duct dilation, IPMN subtype, family history, pancreatitis, unintentional weight loss). In addition, we developed multimodal models integrating MRI sequences in an early fusion manner, using ResNet-34 architecture. Pancreatic region of interest (ROI) was segmented using nnUNet, trained for pancreas segmentation in multiple MRI sequences. Model performance was evaluated through 5-fold cross-validation, and model performance on an independent holdout test set was compared with the clinical gold standard performance based on the Fukuoka consensus guidelines using F1-score and the area under the receiver-operating characteristic curve (AUC). Results: The multimodal model trained on both imaging and clinical data (F1-score: 0.83, 95% CI: 0.61, 1.00; AUC: 0.93, 95% CI: 0.60, 1.00) outperformed the multimodal model trained on imaging data only (F1-score: 0.71, 95% CI: 0.45, 0.91; AUC: 0.87, 95% CI: 0.70, 1.00) as well as unimodal models trained on T2-weighted MRIs alone (F1-score: 0.59, 95% CI: 0.42, 0.81; AUC: 0.70, 95% CI: 0.44, 0.94), T1-weighted MRIs alone (F1-score: 0.58, 95% CI: 0.40, 0.76; AUC: 0.67, 95% CI: 0.39, 0.92), and T1-weighted contrast-enhanced MRIs alone (F1-score: 0.55, 95% CI: 0.37, 0.74; AUC: 0.47, 95% CI: 0.14, 0.80) on the hold-out test set. In addition, the multimodal model outperformed the stratification performance of the Fukuoka criteria (F1-score: 0.54, 95% CI: 0.34, 0.73; AUC: 0.58, 95% CI: 0.31, 0.82) on the holdout test set, showcasing the clinical potential of multimodal AI to personalize management in patients with IPMN. Conclusions: Our findings highlight the added value of multimodal integration, suggesting that multimodal AI models can serve as a valuable diagnostic tool for individualized clinical management of IPMN. External validation is planned to bridge the transition into clinical practice.