Background: Current preoperative assessment faces limitations, including PI-RADS scoring subjectivity and diagnostic uncertainty in distinguishing high-risk prostate cancer from benign and low-risk lesions. To develop an interpretable ensemble learning framework integrating habitat-based radiomics and peritumoral analysis from multiparametric MRI for preoperative high-risk prostate cancer prediction.
Methods: This retrospective, multi-institutional study included 896 patients with suspected prostate lesions and histopathologically confirmed diagnoses across three centers (January 2018-December 2024). Intratumoral habitat analysis used K-means clustering; peritumoral analysis evaluated 1 mm, 3 mm, and 5 mm expansion rings. Feature selection used minimum Redundancy Maximum Relevance (mRMR) and LASSO regression. Models were validated externally with SHAP analysis for interpretability.
Results: The cohort comprised 398 training, 171 internal validation, and 327 external validation patients. The habitat signature achieved superior performance with AUCs of 0.827 (95% CI: 0.768-0.886) and 0.855 (95% CI: 0.795-0.915) in external validation cohorts, significantly outperforming intratumoral signatures (AUCs: 0.774 and 0.629, p < 0.001) and clinical signatures (AUCs: 0.791 and 0.712, p < 0.001). The 3 mm peritumoral signature performed best (AUC: 0.782-0.793). The combined model achieved the highest performance (AUC: 0.860-0.876). SHAP analysis showed ADC-derived features dominated importance, with habitat region H3 contributing > 70% of selected features.
Conclusion: Integrated habitat and peritumoral radiomics provide robust preoperative risk stratification for prostate cancer, with superior performance from ADC-derived habitat features.
Trial registration: Not applicable. This was a retrospective observational study without prospective trial registration.
Keywords: Habitat analysis; Machine learning; Magnetic resonance imaging; Prostate cancer; Radiomics.
© 2026. The Author(s).