Gastritis is defined as inflammation of the gastric mucosa often with erosive changes, and can be sub-grouped into acute versus chronic gastritis. Helicobacter pylori-associated gastritis (HepG) infection has a high baseline prevalence, roughly impacting 50% of patients in high-risk countries. Autoimmune gastritis (AIG) is significantly less common than HepG, and is estimated to be approximately 0.5–2%. Both HepG and AIG increase the risk of gastric adenocarcinoma. The aim of this study is to validate AI-driven models for images acquired through esophagogastroduodenoscopy (EGD) for early accurate diagnosis between AIG and HepG. In this retrospective study, 525 EGD images from two hospitals were collected and the clinical diagnosis of AIG and HepG was made by experienced gastroenterologists. And the deep learning model was constructed based on MobileNet V2. The model was compared with endoscopists in terms of AUC, accuracy, specificity, precision, and recall. AUC, accuracy, precision, recall and specificity were 0.81, 73.6%, 62.5%, 71.4%, and 75%, respectively, for AIG and HePG gastritis. The average performance of the two endoscopists was 0.641, 60.45%, 48.9%, 42.55% and 70.7%. The deep learning model also showed good performance in external independent validation sets. Our study demonstrates the great potential of deep learning methods in identifying early AIG patients and as a tool for clinical AIDS. In the future, further prospective clinical studies are needed to evaluate the clinical applicability of this model.
Yang et al. (Mon,) studied this question.