Computer-Aided Diagnosis System Based on Deep Learning for Diagnosis of Atypical Endometrial Hyperplasia and Endometrial Cancer on Hysteroscopy

Abstract

Hysteroscopy is an effective tool to diagnose and treat abnormal uterine bleeding or uterine cavity abnormalities, especially precancerous or cancerous disorders. In this study, we proposed an Endometrial Cancer (EC) Computer-Aided Diagnosis System (ECCADx) based on deep learning to boost the diagnostic accuracy for recognizing Atypical Endometrial Hyperplasia (AEH) and EC. ECCADx was developed using a training dataset with 49,646 images from 1,237 patients in Maternal and Child Hospital of Hubei Province (MCH) and two test datasets with 7,243 images from 209 patients in MCH, Tongji Hospital (TJH), and The Second Affiliated Hospital of Zhengzhou University (ZZSH). We compared the diagnostic efficiency between the proposed system and eight gynecological endoscopists from two different hospitals (MCH and TJH) using a hospital cross-testing method. The sensitivity, specificity and AUC of ECCADx were 92.8% (95% CI 85.7-100%), 92.5% (95% CI 86.7-98.3%), and 0.965 (95% CI 0. 931-1) on MCH test dataset (internal data), respectively, superior to two gynecological endoscopists and having no significant difference compared with the other two endoscopists at TJH. For TJH/ZZSH test dataset (external data), the sensitivity, specificity and AUC were 75.2% (95% CI 59.5-90.8%), 95.2% (95% CI 91.5-99.0%), and 0.881 (95% CI 0.789-0.967), respectively, superior to three gynecological endoscopists and having no significant difference compared with the other endoscopist at MCH. ECCADx demonstrated excellent performance in identifying AEH and EC in test datasets from different medical centers. The effectiveness of ECCADx was comparable or even better than those of experienced gynecological endoscopists.