Automated and real-time validation of gastroesophageal varices under esophagogastroduodenoscopy using a deep convolutional neural network: a multicenter retrospective study (with video)

Background and Aims Rupture of gastroesophageal varices is the most common fatal adverse event of cirrhosis. EGD is considered the criterion standard for diagnosis and risk stratification of gastroesophageal variceal bleeding. The aim of this study was to train and validate a real-time deep convolutional neural network (DCNN) system, named ENDOANGEL, for diagnosing gastroesophageal varices and predicting the risk of rupture. Methods After training with 8566 images of endoscopic gastroesophageal varices from 3021 patients and 6152 images of normal esophagus/stomach from 3168 patients, ENDOANGEL was also tested with independent images and videos. It was also compared with endoscopists in several aspects. Results ENDOANGEL, in contrast with endoscopists, displayed higher accuracy of 97.00% and 92.00% in terms of detecting esophageal varices (EVs) and gastric varices (GVs) in an image contest (97.00% vs 93.94% , P < .01; 92.00% vs 84.43%, P < .05). It also surpassed endoscopists for red color signs of EVs and red spots of GVs (84.21% vs 73.45%, P < .01; 85.26% vs 77.52%, P < .05). Moreover, ENDOANGEL achieved comparable performance in the determination of size, form, color, and bleeding signs. ENDOANGEL also had good performance in making treatment suggestions. With regard to predicting risk factors in multicenter videos, ENDOANGEL showed great stability. Conclusions Our data suggest that DCNNs were precise in detecting both EVs and GVs and performed excellently in uncovering the endoscopic risk factors of gastroesophageal variceal bleeding. Thus, the application of DCNNs will assist endoscopists in evaluating gastroesophageal varices more objectively and precisely. (Clinical trial registration number: ChiCTR1900023970.) Rupture of gastroesophageal varices is the most common fatal adverse event of cirrhosis. EGD is considered the criterion standard for diagnosis and risk stratification of gastroesophageal variceal bleeding. The aim of this study was to train and validate a real-time deep convolutional neural network (DCNN) system, named ENDOANGEL, for diagnosing gastroesophageal varices and predicting the risk of rupture. After training with 8566 images of endoscopic gastroesophageal varices from 3021 patients and 6152 images of normal esophagus/stomach from 3168 patients, ENDOANGEL was also tested with independent images and videos. It was also compared with endoscopists in several aspects. ENDOANGEL, in contrast with endoscopists, displayed higher accuracy of 97.00% and 92.00% in terms of detecting esophageal varices (EVs) and gastric varices (GVs) in an image contest (97.00% vs 93.94% , P < .01; 92.00% vs 84.43%, P < .05). It also surpassed endoscopists for red color signs of EVs and red spots of GVs (84.21% vs 73.45%, P < .01; 85.26% vs 77.52%, P < .05). Moreover, ENDOANGEL achieved comparable performance in the determination of size, form, color, and bleeding signs. ENDOANGEL also had good performance in making treatment suggestions. With regard to predicting risk factors in multicenter videos, ENDOANGEL showed great stability. Our data suggest that DCNNs were precise in detecting both EVs and GVs and performed excellently in uncovering the endoscopic risk factors of gastroesophageal variceal bleeding. Thus, the application of DCNNs will assist endoscopists in evaluating gastroesophageal varices more objectively and precisely. (Clinical trial registration number: ChiCTR1900023970.)