Preoperative dual-ROI mpMRI radiomics integrating intratumoral and periprostatic adipose tissue features with clinical factors for predicting positive surgical margins after radical prostatectomy

Objective To develop and validate a preoperative mpMRI-based radiomics–clinical nomogram integrating dual-region radiomics features from intratumoral lesions and periprostatic adipose tissue (PPAT) with clinical factors for predicting positive surgical margins (PSM) after radical prostatectomy (RP) in non-metastatic prostate cancer (PCa). Materials and methods This two-center retrospective study included 423 patients (age, 69.78 ± 6.30 years) with non-metastatic PCa who underwent mpMRI followed by RP between January 2019 and December 2024. Intratumoral and PPAT ROIs were manually segmented on T2-weighted images and apparent diffusion coefficient (ADC) maps. After reproducibility filtering, high-correlation removal, and least absolute shrinkage and selection operator (LASSO) regression, three radiomics signatures were constructed. Independent predictors were identified by multivariable logistic regression and integrated into a radiomics–clinical nomogram. Performance was assessed using ROC/AUC, FDR-adjusted DeLong tests, calibration curves, Hosmer–Lemeshow tests, decision curve analysis (DCA), and 1,000-iteration bootstrap validation; subgroup ROC analyses were also performed. Results Twelve intratumoral and nine PPAT radiomics features were selected, and the combined tumor–PPAT radiomics signature (TPR signature) demonstrated superior predictive performance compared with either single-ROI signature across cohorts (training AUC = 0.824; 95% CI, 0.776–0.865; validation AUC = 0.816; 95% CI, 0.737–0.935). Independent predictors included initial PSA, biopsy ISUP grade group, clinical T stage, and the TPR signature. The radiomics–clinical nomogram showed strong discrimination (AUC = 0.908 and 0.885 in the training and validation cohorts, respectively), with good calibration (Hosmer–Lemeshow p = 0.848 and 0.782) and favorable decision-curve performance across threshold probabilities of 0.1–1.0. Bootstrap resampling confirmed model stability, yielding mean AUCs of 0.908 (95% CI, 0.871–0.939) in the training cohort and 0.888 (95% CI, 0.827–0.936) in the validation cohort. Conclusions A dual-ROI mpMRI radiomics–clinical nomogram integrating intratumoral and PPAT phenotypes enables accurate preoperative prediction of PSM after RP and may aid individualized surgical planning.