Preoperative prediction of positive surgical margins in prostate cancer using multimodal deep learning model: a multicenter study.

PURPOSE: This study aimed to develop a deep learning model based on magnetic resonance imaging (MRI) and clinical features for predicting PSM risk after radical prostatectomy (RP). METHODS: This retrospective multicenter study included 1177 prostate cancer patients who underwent preoperative MRI and RP across eight institutions. A total of 1022 patients from two institutions were used for model training, while 155 patients from six independent centers formed the external validation cohort. A feature disentanglement-based deep learning model (DESM) was developed to isolate disease-specific features from hospital-specific variations. A multimodal fusion model (MDESM) was further constructed by integrating the DESM-derived imaging signature with clinical variables to enhance prediction accuracy and generalizability. Gradient-weighted class activation mapping was applied to provide interpretability by highlighting model attention regions. RESULTS: In the external validation cohort, MDESM achieved an area under the receiver operating characteristic curve of 0.843 (95% CI, 0.770-0.911), significantly outperforming the DESM (0.711, 95% CI, 0.607-0.800, p = 0.003, Z = 2.936) and the clinical-only model (0.676, 95% CI, 0.577-0.770, p < 0.001, Z = 3.420). Decision curve analysis demonstrated the highest net benefit for MDESM across a range of threshold probabilities. CONCLUSION: The MDESM model, based on a feature disentanglement and multimodal fusion strategy, demonstrates the potential to effectively combine MRI and clinical data to achieve accurate PSM prediction. This approach offers a promising tool for preoperative risk stratification and surgical planning in prostate cancer.