An interpretable deep learning model for predicting endometrial cancer molecular subtypes from H&E-stained slides

The molecular subtype of endometrial cancer is important for predicting prognosis and treatment effectiveness. This study aimed to develop an interpretable deep learning model based on H&E-stained whole slide images (WSIs) to predict the molecular subtype of endometrial cancer. Data from the Fudan cohort (n = 364) were used to train an end-to-end prediction network for identifying four molecular subtypes. Two external cohorts-the TCGA (n = 296) and Suzhou (n = 36)-were used to validate model generalizability and potential clinical applicability. We further assessed the correlation between histological and molecular features at both the macro- (WSI) and micro- (patch) levels. The network achieved a macro-average area under the receiver operating characteristic curve (AUROC) of 0.867 (95% CI: 0.823-0.911) in 5-fold cross-validation. The class-wise AUROCs were 0.846 (95% CI: 0.798-0.894) for the microsatellite instability-high (MSI-H) subtype, 0.876 (95% CI: 0.831-0.921) for the nonspecific molecular profile (NSMP) subtype, 0.910 (95% CI: 0.818-1.000) for the p53-abnormal (p53abn) subtype, and 0.835 (95% CI: 0.784-0.886) for the POLE-mutated (POLEmut) subtype. Morphological analysis revealed that MSI-H-subtype tumors exhibited increased stromal lymphocytic infiltration; POLEmut tumors showed higher heterogeneity, solid growth patterns, and elevated tumor grade; p53abn tumors were characterized by papillary growth and serous-like features; while NSMP tumors demonstrated high stromal cellularity. This method provides an accurate and interpretable tool for molecular subtype prediction, offering a theoretical basis for future individualized treatment strategies in endometrial cancer.