Integrating biparametric MRI radiomics with clinical variables improves pre-treatment prediction of prostate cancer recurrence.

Introduction: Radiomics can quantify intratumoral heterogeneity on MRI, providing complementary information beyond clinical predictors. This study aimed to evaluate whether integrating radiomic features from pre-operative biparametric MRI (bpMRI) with standard clinical variables improves pre-treatment prediction of biochemical recurrence after radical prostatectomy. We further assessed whether radiomic features add prognostic value to established predictors and compared model performance against the D'Amico classification. Methods: In this retrospective single-center study (2015-2023), 395 men who underwent pre-operative bpMRI (3T Siemens Magnetom Skyra) before radical prostatectomy were included. Lesions were automatically detected using the in-house developed PROVIZ framework, and radiomic features were extracted from index lesions using PyRadiomics v3.1.0. A total of 153 features, including first-order, textural, shape, and anatomical descriptors, were derived from T2-weighted (T2W), ADC, and high b-value diffusion-weighted (DWI, b=1500 s/mm²) images. Clinical variables included prostate specific antigen (PSA), Gleason Grade Group (GGG), PI-RADS v2.1, clinical T stage, and age. A stacked ensemble model (Random Forest and regularized Logistic Regression as base model; Logistic Regression as meta-model) was developed using five-fold stratified cross-validation, SMOTE balancing, Optuna hyperparameter tuning, and isotonic regression-based probability calibration. Performance was evaluated using AUC, calibration, and decision-curve analysis (DCA). Prognostic value was assessed with Kaplan-Meier and Cox regression analyses. Results: The combined model achieved an AUC of 0.85 (95% CI 0.83-0.87), outperforming radiomics-only (0.78) and clinical-only (0.72) models. Calibration was strong (slope = 1.01; Brier = 0.13). DCA showed higher net benefit than D'Amico classification. High-risk patients (probability ≥ 0.26) had significantly shorter recurrence-free survival (log-rank p<0.001; HR = 5.03, 95% CI 2.7-9.9). The most influential predictors were Gleason Grade Group and PSA, together with radiomic first-order/texture features. Conclusion: Integrating bpMRI-derived radiomic features with standard clinical variables improved pre-treatment prediction of biochemical recurrence after radical prostatectomy. The combined model provided better discrimination between high-risk and low-risk recurrence groups and showed higher clinical net benefit than D'Amico. The results support the potential role of radiomics for refining individualized recurrence risk assessment in prostate cancer. External validation in independent cohorts is required before clinical implementation.