Diagnosis subtype of endometrial carcinoma using a deep learning model based on histopathological images

Abstract The histological determination of endometrial carcinoma (EC) remains the keystone in the diagnostic process and helps patients receive appropriate and effective treatment. We focused on constructing a prognostic deep learning (DL) model, which was trained and validated using four independent cohorts of 4108 Hematoxylin and Eosin (H&E) stained images from 1027 patients with multiple magnifications. The model yields an averaged area under the receiver operating curve (AUROC) of 0.962 from the internal dataset, an AUROC of 0.878 on an external dataset from another medical center, and an AUROC of 0.930 on the different scanners dataset, and an AUROC of 0.926 on the poor quality slides dataset, respectively. The predictive value by DL was significant with the mutant P53 of EC (p<0.05). Furthermore, the predictive performance of the model was better than the primary and senior pathologist's prediction (AUC= 0.962, 0.722 and 0.868, respectively). The DL model can accurately highlight morphological characteristics in H&E images and provide indications into the diagnosis and stratification for the management.