This retrospective study enrolled 137 prostate cancer patients who underwent radical prostatectomy at the First Affiliated Hospital of Soochow University Hospital between January 2021 and April 2023. Inclusion criteria (1): Preoperative biparametric MRI including ADC and T2WI sequences (2); Surgery within 4 weeks post-MRI (3); Complete clinicopathological data (age, PSA, biopsy Gleason grade, prostate volume). Exclusion criteria (1): Poor image quality (motion artifacts or incomplete coverage) (2); Prior prostate surgery or hormonal therapy. Patients were classified into adverse-pathology-positive (n=85, ≥1 feature: extracapsular extension, positive surgical margins, lymphovascular invasion) and negative (n=52) groups. Baseline characteristics (age, PSA) were balanced between groups (p>0.05). The study protocol was approved by the First Affiliated Hospital of Soochow University Ethics Committee, with waived informed consent due to retrospective anonymized data.

MRI scans were performed using Siemens Skyra 3.0T and GE Signa 3.0T systems with standardized protocols ( Table 1 ). To mitigate scanner variability, ADC maps were normalized using ComBat harmonization. Two radiologists (5+ years’ experience) independently delineated tumor ROIs on ADC maps, referencing T2WI and DWI (b=50,1000 s/mm²). Inter-observer agreement was assessed by Dice similarity coefficient (mean=0.82). ROIs were rigidly registered across sequences using 3D-Slicer v5.3.0 and resampled to isotropic voxels (1×1×1 mm³).

A total of 851 radiomics features (shape, first-order, texture, wavelet) were extracted via PyRadiomics. To ensure reproducibility, 30 randomly selected cases were re-annotated by the same observer after 2 months, retaining features with ICC≥0.75. Z-score normalization was applied to all features.

The dataset was randomly split into training (n=95) and validation (n=42) sets (7:3 ratio). In the training set:

A clinical model was built using three variables selected by AIC-based stepwise regression: biopsy Gleason grade, total PSA, and prostate volume. The top-performing radiomics model (random forest) was integrated with clinical features to construct a hybrid model.

Model performance was evaluated by AUC, sensitivity, specificity, and accuracy. Calibration curves (Brier score) and decision curve analysis (net benefit threshold: 10–30%) assessed clinical utility. Statistical significance between models was tested via DeLong’s test (AUC comparison) and bootstrapping (1000 iterations). Analyses were conducted in R v4.3.1 (glmnet, caret, pROC packages).

A total of 137 prostate cancer patients were enrolled, including 85 with adverse pathological features (positive group) and 52 without (negative group). The cohort was randomly divided into training (n=95, 59 positive, 36 negative) and validation (n=42, 26 positive, 16 negative) sets at a 7:3 ratio. Baseline clinical characteristics (e.g., age, total PSA, prostate volume) showed no significant differences between training and validation sets (p > 0.05), ensuring balanced group allocation ( Tables 2 and 3 ).

Decision Curve Analysis: The hybrid model provided the highest net benefit across risk a clinically relevant range of threshold probabilities (20–70%), with a 32% reduction in overtreatment observed at a threshold of 0.8 compared to clinical- or radiomics-only strategies ( Figure 3 ).

Nomogram Application: A patient with biopsy Gleason grade 3, PSA=15 ng/mL, and Radscore=1.8 would receive a total risk score of 78%, guiding high-risk classification ( Figure 4 ).

Our study demonstrates that bpMRI-based radiomics models outperform clinical models in predicting APPFs, with the ADC-T2WI combined random forest model achieving a validation AUC of 0.832. This aligns with prior evidence that radiomics captures tumor heterogeneity beyond conventional imaging (22, 23). The superiority of ADC-T2WI over single sequences likely stems from their complementary roles: ADC quantifies cellular density via restricted diffusion (24), while T2WI delineates macrostructural invasion (e.g., capsular irregularity) (25). Notably, our bpMRI model rivals the diagnostic performance of multiparametric MRI (mpMRI) models reported by Gandaglia et al. (AUC=0.81) (26), suggesting that DCE sequences—despite providing hemodynamic data—may offer marginal gains insufficient to justify their added cost and scan time in preoperative prognostication.

Random forest algorithms excelled in our cohort, consistent with Shu et al.’s findings in high-risk prostate cancer stratification (AUC=0.89) (27). This algorithm’s resistance to overfitting and ability to rank feature importance enhance clinical interpretability. In contrast, SVM’s underperformance may reflect our cohort’s limited sample size (n=137) and LASSO-driven linear feature selection, which constrained its capacity to resolve nonlinear boundaries (28).

The clinical model, incorporating biopsy Gleason grade, PSA, and prostate volume, achieved moderate performance (AUC=0.772). While Gleason grade and PSA are established predictors of tumor aggression (29, 30), prostate volume’s inverse correlation with APPFs may arise from PSA dilution in larger glands dominated by benign hyperplasia (31, 32). However, clinical variables alone fail to capture microscale heterogeneity (e.g., focal extracapsular extension), underscoring the need for radiomics integration.

The hybrid model (AUC=0.909) exemplifies the translational potential of multimodal integration. Radiomics features encode tumor microarchitecture (e.g., wavelet textures reflecting stromal fibrosis), while clinical variables contextualize systemic disease burden. This synergy mirrors Fan et al.’s mpMRI-based model (AUC=0.857) (32) but achieves higher accuracy at lower cost—a critical advance for resource-constrained settings. These findings are consistent with emerging evidence that integrating mpMRI radiomics with key clinical variables consistently enhances prognostic performance. For instance, Prata et al. (33) demonstrated superior discrimination of clinically significant prostate cancer when textural features were combined with serum and biopsy data (AUC ≈ 0.80). Similarly, Santucci et al. (34) reported that radiomics-augmented radiomics-RF models outperformed established clinical nomograms for predicting lymph-node involvement (AUC 0.89 vs. 0.79 for the best nomogram).

Clinically, a nomogram-derived risk probability exceeding 80% may serve as a decision threshold for high-risk classification. The hybrid model, implemented as a nomogram, provides individualized risk estimates to support treatment decisions. Patients with lower predicted risk (e.g., <50%) may be eligible for nerve-sparing surgery or active surveillance, while those above 80% are likely high-risk and may need aggressive treatment. As shown in the DCA, applying 80% risk threshold, the hybrid model led to a 32% relative reduction in overtreatment compared to clinical or radiomics-only strategies.

Our study has limitations. First, the single-center retrospective design (n=137) risks selection bias and overfitting, evidenced by the hybrid model’s wide bootstrap AUC CI (0.82–0.97). External validation across diverse populations and MRI platforms is essential. Although random-forest hyper-parameters were chosen by an inner five-fold cross-validation grid search, the AUC nevertheless fell from 0.943 (cross-validated training estimate) to 0.832 in the 42-patient hold-out set. Given the small size of the validation cohort and the residual optimism intrinsic to internal cross-validation, such a decline is expected and underscores the importance of forthcoming multi-institutional external testing. Second, manual ROI delineation, despite high inter-observer agreement (Dice=0.82), introduces subjectivity. Deep learning-based segmentation could improve reproducibility. Third, although we applied intensity normalization and ComBat harmonization, we did not formally quantify radiomics-feature stability across scanner vendors (Siemens vs GE) or across repeated time-points. Scanner-specific hardware, gradient non-linearities and coil configurations can all influence feature distributions, and dedicated phantom or repeat-scan studies will therefore be required to confirm the effectiveness of harmonization in future work. Future work should: 1) Expand data sources: Integrate genomic markers (e.g., PTEN loss) and advanced MRI sequences (e.g., VERDICT) to refine biological specificity. 2) Optimize clinical integration: Develop real-time risk calculators embedded in PACS systems, enabling point-of-care decision support.

In conclusion, this study establishes that bpMRI radiomics-clinical hybrid models predict prostate cancer APPFs with high accuracy (AUC=0.909), suggests potential to reduce reliance on contrast-enhanced imaging in selected patients. By quantifying both microscopic heterogeneity (via radiomics) and macroscopic disease burden (via clinical variables), our approach offers a cost-effective tool for personalized surgical planning. Future multicenter trials should validate these findings and explore AI-driven automation to bridge the gap between radiomics research and clinical implementation.