Deep Learning for Diagnosis of Endometrial Cancer and Atypical Endometrial Hyperplasia

Endometrial cancer(EC) is the most common and rapidly increasing female cancer globally. Atypical endometrial hyperplasia (AEH) is a precancerous condition of EC. Although hysteroscopy serves as the primary modality for diagnosing lesions, it relies on the subjective judgment of hysteroscopists. Therefore, this study proposed a computer-aided diagnostic system utilizing the EfficientNet network as a baseline, incorporating ParNet attention mechanism and class weighting to accurately classify EC/AEH from benign lesions. This study included 49,556 hysteroscopy images from 1,237 cases as a training set and 3,412 hysteroscopy images from 85 cases as a testing set. AUC, accuracy, sensitivity, specificity, PPV, Kappa, and F
1
-Score of the proposed method are 0.941, 89.4%, 93.7%, 87.1%, 73.3%, 0.755, and 0.8225, respectively. The proposed model may be used as a computer-aided tool for the diagnosis of EC/AEH.