A practical AI model for endoscopic diagnosis of gastric premalignant lesions and its application to predicting metachronous recurrence.

BACKGROUND/AIMS: To evaluate the diagnostic performance of an artificial intelligence (AI)-based optical system for chronic atrophic gastritis (CAG) and intestinal metaplasia (IM), and its ability to predict metachronous recurrence. METHODS: We analyzed 72,794 endoscopic images from 5,000 patients undergoing esophagogastroduodenoscopy with biopsy-confirmed CAG/IM. A convolutional neural network model was developed and compared with experts (n=16) and novices (n=15). The association between AI-predicted IM extent and metachronous recurrence after endoscopic resection of early gastric cancer (n=458) was evaluated using Cox regression. RESULTS: For CAG, the AI model achieved accuracies of 0.819 (antrum) and 0.864 (body), with corresponding AUCs of 0.885 and 0.901. For IM, accuracies were 0.759 and 0.723, with AUCs of 0.835 and 0.744. The AI model outperformed novices and experts in CAG diagnosis (accuracy: 0.855 vs. 0.705 and 0.751) and outperformed novices in IM diagnosis (0.875 vs. 0.781). Although individual HRs for AI-predicted IM extent were not statistically significant (2.11 [95% CI, 0.25-17.67], antral IM only; 4.27 [0.57-31.97], IM extending to body), a significant dose-response relationship with recurrence risk was observed (Ptrend=0.043). CONCLUSIONS: This AI model showed clinically acceptable performance and potential usefulness for CAG/IM diagnosis in a single-center study but requires multicenter external validation before clinical adoption.