MRI-based radiomics interpretable model predicts microsatellite instability status in endometrioid adenocarcinoma

To develop and validate a multiparametric MRI-based radiomics model for noninvasive preoperative prediction of microsatellite instability (MSI) in patients with endometrioid adenocarcinoma (EC), and to improve model interpretability with SHapley Additive exPlanations (SHAP) analysis.

We enrolled 241 patients with pathologically confirmed EC and divided them into training (n = 193) and test (n = 48) cohorts at an 8:2 ratio. Radiomic features were extracted from T2-weighted imaging (T2WI), apparent diffusion coefficient (ADC), and late-phase T1-weighted contrast-enhanced (T1CE) sequences. Feature screening was performed sequentially via intraclass correlation coefficient (ICC) filtering, Pearson collinearity elimination, and 10-fold cross-validated least absolute shrinkage and selection operator (LASSO) regression. Five machine learning models were constructed and evaluated via area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. The optimal model was validated using decision curve analysis (DCA), and SHAP analysis was applied to interpret feature contributions.

The MSI rate was 29% (70/241). Four features were retained after selection. The XGBoost model achieved the best performance (training AUC = 0.983, test AUC = 0.870), with test set accuracy of 0.812, sensitivity of 0.929, and specificity of 0.765. DCA confirmed clinical utility, SHAP analysis identified T2WI-GLSZM GrayLevelVariance as the most impactful feature and visualized the differences in feature contributions between MSI and MSS patients at the individual level, clarifying the model's decision-making logic.

The multiparametric MRI-based radiomics model enables noninvasive prediction of MSI in EC, potentially guiding personalized immunotherapy selection.