Interpretable Machine Learning with MRI Habitat Radiomics for Preoperative Assessment of Microsatellite Instability in Rectal Cancer

Motivation: Tumor heterogeneity in rectal cancer limits the detection of microsatellite instability. Goal(s): An interpretable machine learning model based on MRI subregional habitat radiomics and clinical features was constructed to preoperatively evaluate MSI status in rectal cancer patients. Approach: A retrospective analysis of 291 rectal cancer patients was conducted. K-means clustering segmented the tumors, and subregional and classical radiomic features were extracted from multiparametric MRI. Three models were developed based on logistic regression. Shapley Additive Explanation quantified each feature's contribution to model predictions. Results: The combined model demonstrated the best performance, with AUCs of 0.908 and 0.863 in the training and validation cohorts, respectively. Impact: This study developed and validated a combined model using multiparametric MRI subregional radiomics, classical radiomics, and clinical variables for non-invasive preoperative MSI prediction. SHAP aids personalized predictions, supporting individualized treatment and biopsy-targeted decision-making in RC patients.