A rural-to-center artificial intelligence model for diagnosing Helicobacter pylori infection and premalignant gastric conditions using endoscopy images captured in routine practice
Yi‐Chia Lee, Chu-Song Chen, Tsung‐Hsien Chiang, Yen-Ning Hsu, Min-Han Chen, Yi-Ru Chen, Hsiu-Chi Cheng, Mei-Jin Chen, Fu‐Jen Lee, Chi-Yang Chang, Chun‐Chao Chang, Ming‐Jong Bair, Jyh‐Ming Liou, Chiuan-Jung Chen, Yen‐Chung Chen, Hung Chiang, Chia‐Tung Shun, Jui-Hsuan Liu, Han‐Mo Chiu, Ming‐Shiang Wu, Jiun-Yu Yu, Ruey-Shan Guo, Jaw‐Town Lin
Abstract
Background: Diagnosing Helicobacter pylori infection and premalignant gastric conditions typically requires 13C urea breath testing or histological assessment, which are often unavailable in remote areas. A rural-to-center artificial intelligence (AI) model was developed and implemented to automatically evaluate upper endoscopy images from routine clinical practice. Methods: Endoscopic images were collected from a rural hospital on Matsu Islands and a tertiary center across Taiwan Strait. During model development (2020–2022), AI algorithms were trained, validated, and tested to exclude low-quality and non-gastric images, segment gastric regions, and enhance mucosal features for detecting H. pylori infection and premalignant conditions. During model implementation (2023–2024), endoscopic images from a rural hospital were transmitted to the medical center for AI analyses, with results promptly returned. Results: In the development phase, diagnostic accuracies were 92.8% (95%CI 88.9%–96.6%) for H. pylori, 88.6% (95%CI 87.2%–90.0%) for atrophic gastritis, and 88.0% (95%CI 86.5%–89.5%) for intestinal metaplasia. In the implementation phase, 3518 rural residents underwent 13C urea breath testing or pepsinogen testing; 421 with positive results underwent endoscopy. No significant differences were observed between AI-predicted and clinically observed prevalence: H. pylori (13.9% vs. 12.9%; P = 0.55), atrophic gastritis (15.7% vs. 11.9%; P = 0.34), and intestinal metaplasia (27.6% vs. 22.4%; P = 0.32). Implementation-phase diagnostic accuracies were 91.3% (95%CI 88.0%–94.6%), 79.9% (95%CI 72.1%–86.3%), and 63.4% (95%CI 54.7%–71.6%), respectively. Conclusions: AI enabled physicians in resource-limited settings to rapidly assess gastric health using routinely captured endoscopic images, bridging gaps in access and expertise.