Multimodal integration of VI-RADS, radiomics, and deep learning features for prediction of pathology and prognosis in bladder cancer.

640 Background: Recently, MRI-based assessments of bladder cancer using VI-RADS, radiomics, and deep learning models have been widely reported. However, it has not been known whether MRI-based approaches can comprehensively predict detailed pathological findings and survival outcomes. This study aimed to evaluate a multimodal approach integrating VI-RADS, radiomics, and deep learning features for predicting both pathology and survival outcomes. Methods: We retrospectively analyzed 106 patients with bladder cancer who underwent MRI and TURBT between August 2021 and July 2023. VI-RADS were assessed by two urologists. Tumors were manually segmented in 3D slicer, and radiomic features were extracted with PyRadiomics. Deep learning–derived features were obtained from cropped tumor images using EfficientNet-B7. Detailed pathological outcomes were modeled after feature selection with LASSO regression and cross-validation. For survival analysis, OS and PFS at 3 years were evaluated. Risk scores were derived from LASSO-selected features, and cutoff values were determined by ROC analysis to stratify patients into risk groups. Results: The multimodal approach showed favorable predictive performance. In this study, features selected by LASSO consisted of 85% from EfficientNet, 10% from VI-RADS, and 5% from radiomics. For pathological findings, neural boosting achieved high accuracy with AUCs of 0.995 for MIBC, 0.970 for T stage, 0.965 for high grade, 0.983 for necrosis, 0.891 for variant histology, and 0.917 for CIS. For survival outcomes, the risk score achieved AUCs of 0.835 for 3-year OS and 0.926 for 3-year PFS. Kaplan–Meier analysis demonstrated significantly poorer overall survival (HR = 13.99, 95% CI 3.84–50.58) and progression-free survival (HR = 18.79, 95% CI 4.11–85.79) in the high-risk group compared with the low-risk group. Conclusions: Integration of MRI-derived VI-RADS, radiomics, and deep learning features enables accurate prediction of pathology and prognosis in bladder cancer without additional clinical factors. This MRI-based multimodal risk score may support individualized patient stratification and treatment decision-making.