Identification of upper GI diseases during screening gastroscopy using a deep convolutional neural network algorithm

Background and Aims The clinical application of GI endoscopy for the diagnosis of multiple diseases using artificial intelligence (AI) has been limited by its high false-positive rates. There is an unmet need to develop a GI endoscopy AI-assisted diagnosis system (GEADS) to improve diagnostic accuracy and clinical utility. Methods In this retrospective, multicenter study, a convolutional neural network was trained to assess upper GI diseases based on 26,228 endoscopic images from Dazhou Central Hospital that were randomly assigned (3:1:1) to a training dataset, validation dataset, and test dataset, respectively. To validate the model, 6 external independent datasets comprising 51,372 images of upper GI diseases were collected. In addition, 1 prospective dataset comprising 27,975 images was collected. The performance of GEADS was compared with endoscopists with 2 professional degrees of expertise: expert and novice. Eight endoscopists were in the expert group with >5 years of experience, whereas 3 endoscopists were in the novice group with 1 to 5 years of experience. Results The GEADS model achieved an accuracy of .918 (95% confidence interval [CI], .914-.922), with an F1 score of .884 (95% CI, .879-.889), recall of .873 (95% CI, .868-.878), and precision of .890 (95% CI, .885-.895) in the internal validation dataset. In the external validation datasets and 1 prospective validation dataset, the diagnostic accuracy of the GEADS ranged from .841 (95% CI, .834-.848) to .949 (95% CI, .935-.963). With the help of the GEADS, the diagnosing accuracies of novice and expert endoscopists were significantly improved (P < .001). Conclusions The AI system can assist endoscopists in improving the accuracy of diagnosing upper GI diseases.