MRI habitat analysis using interpretable machine learning models for ISUP grade group prediction in prostate cancer

To evaluate the diagnostic performance of a noninvasive preoperative ISUP classification model for prostate cancer based on biparametric MRI (bpMRI) habitat imaging (HI). Retrospective data were collected from patients with pathologically confirmed prostate cancer in two medical centers between July 2021 and August 2024, including clinical information, pathological results, and bpMRI (T2WI + DWI) features. Patients from Center 1 were randomly assigned to the training and internal validation cohorts at a 7:3 ratio, whereas those from Center 2 were adopted as the external validation cohort. The K-means clustering algorithm was applied to segment habitat subregions, followed by radiomic feature extraction from each subregion. Predictive models were established using five machine learning algorithms, and SHapley Additive exPlanations (SHAP) analysis was performed to quantify and visualize feature importance. The predictive performance of all models for prostate cancer ISUP classification was assessed using ROC analysis. Calibration curves and decision curve analysis were further adopted to evaluate model calibration and clinical net benefit. A total of 412 patients were enrolled in this study, comprising a training cohort ( n = 196), an internal validation cohort ( n = 82), and an external validation cohort ( n = 134). Multivariable analysis identified total prostate-specific antigen (tPSA) (OR = 1.041, p < 0.001) and ADC ratio (mean ADC of tumor/normal tissue) (OR < 0.001, p < 0.001) as independent predictors for prostate cancer ISUP classification. Of the 40 established predictive models, the comprehensive model integrating clinical factors, conventional imaging features, and radiomic features derived from T2WI habitat subregion 3 yielded the optimal performance. AUC values of the comprehensive model were 0.867, 0.900, and 0.826 in the training, internal validation, and external validation cohorts, respectively. The model also achieved superior predictive efficacy and clinical net benefit compared to all other models. The bpMRI–based comprehensive model integrating clinical, imaging and habitat radiomic features enables noninvasive preoperative prediction of prostate cancer ISUP classification. It may serve as an auxiliary tool to facilitate individualized preoperative treatment decision-making.