Diagnostic value of whole-tumor ADC histogram parameters combined with ISUP grade and MRI-EPE score for predicting extracapsular extension in PI-RADS 4-5 prostate cancer.

PURPOSE: To investigate the incremental value of whole-tumor apparent diffusion coefficient (ADC) histogram parameters, when combined with clinical pathological variables for predicting extracapsular extension (EPE) in patients with PI-RADS 4-5 prostate cancer. METHODS: This retrospective study included 102 patients with pathologically confirmed prostate cancer following surgery, comprising 51 patients with EPE and 51 without EPE as the training cohort, plus an independent temporal validation cohort of 34 patients. All patients underwent preoperative multiparametric magnetic resonance imaging (mpMRI). Tumor regions and 1-3 mm peritumoral regions of interest (ROIs) were manually segmented on ADC maps using 3D Slicer software, followed by the extraction of histogram parameters. Histogram features were selected using forward stepwise logistic regression based on the likelihood ratio test. Clinical, histogram, and combined predictive models were established using logistic regression analysis. Diagnostic performance was assessed using receiver operating characteristic (ROC) curve analysis. Internal validation was performed with 1000 bootstrap resamples to evaluate model stability and optimism. Calibration and decision curve analyses were further applied to assess model accuracy and clinical utility. To specifically quantify incremental value beyond the established clinical model, continuous net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were defined as core secondary endpoints and calculated in the temporal validation cohort. RESULTS: Multivariate Firth-corrected logistic regression analysis identified the International Society of Urological Pathology (ISUP) grade (OR = 2.701, p = 0.033) and MRI-EPE score (OR = 5.441, p = 0.020) as independent clinical predictors of EPE. Among the single-region histogram models, the peri_1mm model achieved the highest diagnostic performance (AUC = 0.839). Peri1_Minimum and peri3_Minimum were selected as the most informative histogram features for model construction. The combined model, incorporating ISUP grade, MRI-EPE score, peri1_Minimum, and peri3_Minimum, demonstrated the best overall performance with an AUC of 0.917, which was significantly higher than that of the histogram model alone (AUC = 0.858, p = 0.026), while showing no significant difference compared with the clinical model (AUC = 0.903, p = 0.336). Bootstrap internal validation yielded corrected AUC values of 0.904, 0.890, and 0.845 for the combined, clinical, and histogram models, respectively. The combined model also showed excellent calibration and provided greater net clinical benefit across a broader range of threshold probabilities on decision curve analysis. Notably, in the independent temporal validation cohort, although the combined model's AUC (0.861) remained numerically higher than that of the clinical model (0.818, p = 0.208), the addition of peritumoral ADC histogram features resulted in a significant continuous NRI of 0.7286 (95% CI: 0.0971-1.3601, p = 0.024) and a borderline significant IDI of 0.0919 (p = 0.053), indicating that the addition of peritumoral ADC histogram features led to a net improvement in the direction of predicted risk changes, beyond what was provided by the clinical model alone. The combined model also showed good calibration, with a calibration slope of 0.710 (95% CI: 0.273-1.148) and a Brier score of 0.150. CONCLUSION: Integration of ISUP grade, MRI-EPE score, and peritumoral ADC histogram features (peri1_Minimum and peri3_Minimum) enables robust prediction of EPE in PI-RADS 4-5 prostate cancer. While the combined model's overall discriminatory performance was comparable to that of the clinical model based on AUC, the significant continuous NRI suggests incremental value in refining risk stratification. However, given the non-significant AUC difference and borderline IDI, these findings warrant validation in larger, multicenter cohorts before informing individualized surgical planning for high-risk patients.