Multiparametric MRI radiomics combined with clinical-radiologic features improves preoperative prediction of muscle invasion in bladder cancer

Abstract

Accurate preoperative differentiation of muscle-invasive bladder cancer (MIBC) from non-muscle-invasive bladder cancer (NMIBC) is central to treatment selection. We evaluated whether a nomogram integrating multiparametric MRI (mpMRI) radiomics, VI-RADS, and routinely recorded clinical-radiologic features improves prediction of muscle invasion. This single-center retrospective study included 350 consecutive patients with histologically confirmed bladder urothelial carcinoma who underwent preoperative 3.0-T mpMRI between January 2019 and December 2024. Patients were split chronologically into a training cohort (n = 245) and a temporal validation cohort (n = 105). Radiomics features were extracted from T2-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient maps. Reproducibility filtering, univariable screening, collinearity pruning, and least-absolute-shrinkage-and-selection-operator regression were performed using the training cohort only. Clinical-radiologic, radiomics, and combined logistic models were evaluated by discrimination, calibration, Brier score, and decision curve analysis. MIBC was present in 78 of 245 training patients (31.8%) and 33 of 105 validation patients (31.4%). Rad-score, VI-RADS score, tumor size, hydronephrosis, and absence of a tumor stalk remained independently associated with muscle invasion. The combined nomogram achieved AUCs of 0.929 (95% CI 0.895-0.963) in training and 0.897 (95% CI 0.839-0.955) in validation, compared with 0.708 for the clinical-radiologic model, 0.798 for VI-RADS, and 0.838 for the radiomics model in validation. Among the 59 patients with a research VI-RADS score of exactly 3, the combined nomogram achieved AUCs of 0.871 (95% CI 0.751-0.991) in training and 0.835 (95% CI 0.625-0.999) in validation; the pooled descriptive AUC was 0.864 (95% CI 0.764-0.964). Because VI-RADS was constant within this subgroup, an informative VI-RADS-alone ROC curve could not be estimated. Validation calibration intercept and slope for the combined model were 0.06 (95% CI -0.20 to 0.32) and 0.91 (95% CI 0.70-1.13), respectively. A combined mpMRI radiomics-clinical-radiologic nomogram showed strong temporal-validation performance for preoperative MIBC prediction, with an exploratory signal of utility in VI-RADS 3 lesions. External multicenter validation is required before clinical implementation.