A deep learning-based computer-aided diagnosis system for detecting atypical endometrial hyperplasia and endometrial cancer through hysteroscopy.

Timely diagnosis of endometrial cancer (EC) and atypical endometrial hyperplasia (AEH) is crucial, yet traditional hysteroscopy faces accuracy challenges. This study introduces ECCADx, a deep learning-based computer-aided diagnosis system utilizing contrastive learning for hysteroscopic identification of AEH and EC. This is the system to integrate contrastive learning for this specific differentiation. ECCADx leveraged contrastive learning during pre-training on diverse external medical images, extracting robust features. Trained on 49,646 images from 1,204 patients, it underwent rigorous multicenter validation on two independent test datasets (6,228 images from 190 patients). ECCADx consistently achieved high diagnostic accuracy, often surpassing experienced endoscopists. Notably, it attained 95.2% sensitivity and 91.3% specificity on the internal dataset, and 92.1% sensitivity with 100% specificity on the external dataset. ECCADx proves a reliable tool, comparable or superior to human experts, promising to reduce misdiagnosis and improve patient outcomes.