Interpretable machine learning model for predicting clinically significant prostate cancer: integrating intratumoral and peritumoral radiomics with clinical and metabolic features.

Background: To develop and validate an interpretable machine learning model based on intratumoral and peritumoral radiomics combined with clinicoradiological features and metabolic information from magnetic resonance spectroscopy (MRS), to predict clinically significant prostate cancer (csPCa, Gleason score ≥ 3 + 4) and avoid unnecessary biopsies.

Methods: This study retrospectively analyzed 350 patients with suspicious prostate lesions from our institution who underwent 3.0 Tesla multiparametric magnetic resonance imaging (mpMRI) prior to biopsy (training set, n = 191, testing set, n = 83, and a temporal validation set, n = 76). Intratumoral and peritumoral volumes of interest (VOI, VOI)) were manually segmented by experienced radiologists on T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC) maps. Radiomic features were extracted separately from the VOI and VOI. After feature selection via the recursive feature elimination (RFE) algorithm, intratumoral radiomic score (intra-rad-score) and peritumoral radiomic score (peri-rad-score) were constructed. The clinical model, MRS model, and combined model integrating radiomic, clinicoradiological and metabolic features were constructed via the eXtreme Gradient Boosting (XGBoost) algorithm. The predictive performance of the models was evaluated in both the training and testing sets using receiver operating characteristic (ROC) curve analysis. SHapley Additive exPlanations (SHAP) analysis was applied to the combined model to visualize and interpret the prediction process.

Results: A total of 350 patients were included, comprising 173 patients with csPCa (49.4%) and 177 patients with non-csPCa (50.6%). The intra-rad-score and peri-rad-score were constructed via 10 and 16 radiomic features. The combined model demonstrated the highest AUC, accuracy, F1 score, sensitivity, and specificity in the testing set (0.968, 0.928, 0.927, 0.932, and 0.923, respectively) and in the temporal validation set (0.940, 0.895, 0.890, 0.923, and 0.875, respectively). SHAP analysis revealed that the intra-rad-score, PSAD, peri-rad-score, and PI-RADS score were the most important predictors of the combined model.

Conclusion: We developed and validated a robust machine learning model incorporating intratumoral and peritumoral radiomic features, along with clinicoradiological and metabolic parameters, to accurately identify csPCa. The prediction process was visualized via SHAP analysis to facilitate clinical decision- making.