Abstract / Summary
Abstract Background This study developed and validated a machine-learning model using biparametric MRI (bpMRI) that integrates clinical variables, Prostate Imaging Reporting and Data System (PI-RADS) scores, and radiomics features to distinguish clinically significant prostate cancer (csPCa). Methods This retrospective study included 458 patients with clinical and bpMRI data. Clinical predictors were identified via univariate and multivariate stepwise regression. Regions of interest (ROIs) were delineated on T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC) maps using 3D Slicer, and radiomics features were extracted with PyRadiomics. Optimal features were selected via dimensionality reduction and least absolute shrinkage and selection operator (LASSO) regression. Patients were split 7:3 into training and validation sets. LR, SVM, RF, and XGBoost were evaluated as candidate algorithms for the radiomics and combined models. The final algorithm was selected based on discrimination, calibration, and clinical utility. Model performance was assessed by discrimination (area under the receiver operating characteristic curve [AUC]), calibration (calibration slope and intercept), and clinical utility (decision curve analysis within the 10%–60% threshold range). The selected model was then used to build a nomogram. Results In the validation cohort, the clinical model, Radiomics LR, and Combined LR achieved AUCs of 0.843, 0.888, and 0.900, respectively. Combined LR significantly outperformed both the clinical model ( p < 0.001) and Radiomics LR ( p = 0.032). Combined LR showed a sensitivity of 0.841 and a calibration slope of 0.950 (intercept = −0.759). After balancing discrimination, calibration, and clinical utility, Combined LR was selected to build a nomogram, which demonstrated good calibration and net benefit within the 10%–60% threshold range. Conclusion This nomogram integrates bpMRI radiomics, clinical variables, and PI-RADS scores and outperforms conventional radiomics and clinical models. It provides a non-invasive tool for risk stratification before biopsy and helps to reduce unnecessary biopsies.