Deep Learning-Based MRI Segmentation for Non-Invasive Prediction of Microsatellite Instability in Endometrial Cancer: A Multicenter Study

Objective Preoperative determination of microsatellite instability (MSI) status in endometrial cancer (EC) is crucial for guiding immunotherapy, but currently relies on invasive tissue sampling. This study aimed to develop and validate a fully automated, non-invasive framework using deep learning-based MRI segmentation and radiomic analysis for preoperative prediction of MSI status in EC. Methods In this retrospective multicenter study, 867 EC patients from three institutions were enrolled (593 for model development, 274 for external testing). MSI status was determined by immunohistochemistry. A cascaded V-Net model was developed for automated tumor segmentation on T2-weighted, diffusion-weighted, and contrast-enhanced T1-weighted images. Radiomic features were extracted and selected based on interobserver reliability, redundancy removal, and least absolute shrinkage and selection operator regression. Three classifiers—support vector machine (SVM), random forest (RF), and logistic regression (LR)—were trained and externally validated. Segmentation performance was assessed using the Dice similarity coefficient (DSC), sensitivity, and specificity. Classification performance was evaluated by the area under the curve (AUC), accuracy, sensitivity, and specificity, with group comparisons using DeLong tests. Results In the external testing cohort, automated segmentation achieved mean DSCs of 85.8%±13.3% for DWI, 76.2%±17.3% for T2WI, and 80.5%±14.6% for CE-T1WI. The radiomic models based on automated segmentation yielded AUCs of 0.876 (SVM), 0.867 (RF), and 0.800 (LR) for predicting MSI status, which were not significantly different from the performance of models based on manual segmentation (AUCs: 0.915, 0.890, 0.821; all P > 0.05). Conclusion Deep learning-based automated MRI segmentation can reliably delineate EC tumors, and when combined with radiomic analysis, provides an effective, non-invasive method for preoperative prediction of MSI status. This automated framework demonstrates performance comparable to manual segmentation, supporting its potential integration into clinical workflows to aid preoperative decision-making.