DEEP LEARNING ALGORITHM FOR MOLECULAR CLASSIFICATION OF ENDOMETRIAL CANCER FROM WHOLE SLIDE HISTOPATHOLOGY IMAGES

Endometrial cancer (EC) is a common malignancy whose molecular classification (POLEmut, MMRd, p53-abn, NSMP) guides prognosis and treatment. While MMRd and p53-abn can be assessed through IHC, POLEmut identification requires gene sequencing, which is costly and often unavailable. In this study, we developed a fully supervised deep learning (DL) model to classify EC molecular subtypes directly from H&E-stained whole-slide images (WSIs). From an initial cohort of 1,362 cases, 230 FFPE WSIs were selected and annotated to train three sequential binary classifiers (POLEmut vs non-POLE, MMRd vs non-MMRd, p53-abn vs NSMP), forming a hierarchical, clinically aligned architecture. Prediction heatmaps were generated to enhance interpretability. The model showed excellent performance for POLEmut (AUROC 0.95; accuracy 87.5%; F1 0.86) and good performance for MMRd (AUROC 0.88; F1 0.81) and p53-abn (accuracy 74%; F1 0.70). Overall, it achieved an average precision of 76% and recall of 88%. These results demonstrate the feasibility of DL-based prediction of EC molecular subtypes from routine histology, offering a scalable approach to support diagnostic workflows and expand access to precision oncology where molecular assays are limited.