July 28, 2025

Interpretable Machine Learning Model Based on Intra- and Peritumoral MRI Radiomics for Predicting Biochemical Recurrence After Radical Prostatectomy

Key Points

The combined MRI radiomics-clinical model effectively predicts biochemical recurrence post-prostatectomy.
SHAP analysis identifies extraprostatic extension as the primary risk factor for biochemical recurrence.
IntraPeri_4mm model yielded AUC of 0.902, outperforming other models in prediction accuracy.
Machine learning enhances risk stratification and personalized treatment planning for prostate cancer patients.

Abstract

Abstract PurposeTo develop a predictive model for biochemical recurrence (BCR) after radical prostatectomy (RP) by integrating intratumoral and peritumoral MRI radiomics features with clinical independent risk factors. MethodsThis retrospective study analyzed 277 RP patients with complete follow-up data (≥1 year) from our institution, randomly divided into training (n=193) and test (n=84) cohorts. Regions of interest (ROIs) were manually delineated on T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC) maps. Peritumoral ROIs were expanded by 4 mm using Python and manually adjusted to exclude non-prostatic tissues. Radiomics models (Intra, Peri₄mm, IntraPeri₄mm), a clinical model (Clinic), and a combined radiomics-clinical model (Combined) were constructed. The predictive performance of these models was evaluated using different indexes. SHapley Additive exPlanations (SHAP) analysis was employed to visualize and interpret the decision-making process. Results Independent risk factors for BCR included extraprostatic extension (EPE), clinical N stage (N), seminal vesicle invasion (SVI), PI-RADS score, neutrophil-to-lymphocyte ratio (NLR), and maximum transverse diameter of prostate (MTD). The Clinic model achieved AUC of 0. 897 (training) and 0. 731 (test). The IntraPeri₄mmADC model showed AUC of 0. 902 and 0. 706, while the IntraPeri₄mmT2 model yielded 0. 842 and 0. 662. The Combined model outperformed others (AUC: 0. 978 and 0. 810). DCA confirmed its higher net benefit. SHAP analysis revealed EPE as the top contributor to BCR prediction, followed by the ADC-derived radiomics score (ADC label₀). Conclusions The combined MRI radiomics-clinical model effectively predicts BCR post-RP. SHAP interpretability transforms "black-box" predictions into quantifiable feature contributions, aiding clinicians in risk stratification and personalized treatment planning.

问 AI

Bookmark