Real-Time Deep Convolutional Neural Networks Diagnose Esophageal Varices And Record Five Endoscopic Findings

Mingkai Chen,Jing Wang,Yong Xiao,Lianlian Wu,Hong-Gang Yu

DOI: https://doi.org/10.1016/J.GIE.2020.03.1069

IF: 10.396

2020-01-01

Gastrointestinal Endoscopy

Abstract:Gastroesophageal varices are the second complications of cirrhosis and the mortality of varices hemorrhage (VH) is high. Esophagogastroduodenoscopy (EGD) is an essential measure to diagnose gastroesophageal varices and make risk stratification. The aim of this study was to train and validate real-time deep convolutional neutral networks (DCNNs) to diagnose esophageal varices and to predict the risk of rupture by recording five endoscopic findings. 6734 endoscopic esophageal varices images and 3921 normal esophagus images were used to train, validate and test DCNNs called “ENDOANGEL”. ENDOANGEL was tested with independent images and compared the performance with the endoscopists. We also tested ENDOANGEL in a video. Compared to endoscopists, ENDOANGEL showed high diagnostic accuracy of 96.90%. The accuracy of endoscopists for red color signs and mucosal findings categories were significantly lower than that of ENDOANGEL. No significant differences were found between ENDOANGEL and endoscopists in form, color and bleeding signs. In the gastroesophageal video, ENDOANGEL diagnosed and predicted five endoscopic findings in real time. Typical images indicating endoscopic findings were collected and listed in a final report after the examination. This study suggested that ENDOANGEL are capable to diagnose esophageal varices with a high accuracy. EDOANGEL performed well in recording five endoscopic findings. The application to the video indicated that ENDOANGEL could not only increase confidence level of endoscopists for diagnosing esophageal varices but also increase the accuracy of recording five endoscopic findings in a real-time process.View Large Image Figure ViewerDownload Hi-res image Download (PPT)