Abstract
Accurate molecular subtyping is essential for guiding precision treatment and prognostic stratification in endometrial cancer (EC). However, current methods, based on Sanger sequencing and immunohistochemistry (IHC), are costly, time-intensive, and difficult to implement widely in routine clinical practice, particularly in resource-limited settings. To overcome these challenges, we developed a deep-learning pipeline that directly infers EC molecular subtypes from routine hematoxylin-and-eosin (H&E) whole-slide images (WSIs). The framework integrates super-resolution enhancement (SRResGAN), transformer-based lesion segmentation (MedSAM), and a ResNet-101 classifier for molecular subtype prediction, with an LSTM module for survival modeling. This retrospective study included 393 Chinese patients diagnosed between 2010 and 2018, all with ≥ 5 years of follow-up. Molecular subtypes-POLE(mut), mismatch repair-deficient (MMRd), p53abnormal (p53abn), and no specific molecular profile (NSMP)-were confirmed by Sanger sequencing and immunohistochemistry. The model achieved high classification accuracies (92% for POLE(mut) and MMRd, 91% for p53abn, and 90% for NSMP), with a strong correlation between predicted and observed survival (R2 = 0.9692; MAE = 123 days). External validation on two independent cohorts (N = 35 and N = 83) confirmed robust generalizability across institutions. This study represents the first large-scale, multicenter, AI-based digital pathology model for EC molecular classification in China. The proposed workflow provides an automated, interpretable, and cost-efficient alternative to conventional molecular testing, supporting precision oncology, fertility-preserving management, and clinical decision-making in real-world practice.