Gastric cancer remains a major contributor to global cancer mortality. Although endoscopy is the most effective method for detecting early gastric cancer (EGC), the identification of subtle early lesions continues to pose a clinical challenge. Artificial intelligence (AI) has emerged as a promising adjunct to enhance diagnostic accuracy. AI systems are trained using different methodological approaches, and their diagnostic performance may vary. This meta-analysis aimed to evaluate the diagnostic accuracy of AI-assisted endoscopy for EGC and to compare the performance of different AI systems

PubMed, Embase, Cochrane, and Web of Science were systematically searched for articles. Pooled sensitivity and specificity were calculated using random-effects models; an SROC curve was constructed and AUC estimated. Histopathology served as the reference standard. Subgroup analyses were performed according to AI system type and endoscopic imaging modality. Risk of bias was assessed using the QUADAS-2 tool, and publication bias was evaluated with funnel plots

Twenty-one studies were included, conducted in China, Japan and Korea. AI achieved an overall AUC of 0.96 (95% CI, 0.94–0.98), pooled sensitivity of 91% (95% CI, 88–94), and specificity of 90% (95% CI, 85–94), with substantial heterogeneity. Deep-learning systems showed sensitivity comparable to hybrid machine learning–deep learning models (89.9% [95% CI, 84.1–93.8] vs 91.2% [95% CI, 85.5–94.8]) but achieved significantly higher specificity (89.4% [95% CI, 86.3–91.9] vs 83% [95% CI, 79.2–86.3], p ≤ 0.005). No significant methodological or publication bias was identified

AI-assisted endoscopy demonstrates high diagnostic accuracy for the early detection of gastric cancer. AI systems are not equivalent: deep-learning models provide the best performance, particularly due to their superior specificity