An evaluation of machine learning techniques for the diagnosis of gastric cancer: A meta analysis.

e16078 Background: Gastric cancer is a leading cause of cancer mortality worldwide, with early detection crucial for better survival. Recent advancements in Artificial Intelligence (AI) have shown promise in improving diagnostic accuracy across various tested imaging modalities. By modifying various machine learning algorithms, AI models are capable of identifying diverse mucosal patterns that may improve early detection of gastric cancers. This review aims to evaluate the overall diagnostic accuracy of AI-based models in detecting gastric cancers via endoscopy and analyze the diagnostic accuracy of detecting gastric cancers using narrow-band imaging compared to white light imaging. Methods: A systematic search of PubMed and Scopus identified studies evaluating the diagnostic performance of various AI models for gastric cancer based on endoscopic imaging. The statistical analysis for the Meta Analysis was carried out in the R-Studio Package. Random intercept logistic regression model was used to perform the analysis. Pooled specificity and pooled sensitivity were analyzed. A subgroup analysis was conducted to compare white light imaging and arrow band imaging. Additionally, another subgroup analysis was conducted based on artificial intelligence models assessed in the studies. Results: A total of 12 studies, 16 subgroups and 30,646 images were assessed in the meta-analysis. The pooled sensitivity and specificity were estimated to be 86.70% [79.20; 91.80, 95% CI, p < 0.01, I^2 = 98.0% [97.5%; 98.4%] ] and 92.16 [87.85; 95.02, 95% CI, p < 0.01, I^2 = 97.3% [96.5%; 97.9%] utilising the random effects model. High sensitivity and specificity were observed in Region-based CNN models in comparison to the conventional CNN models. In the analysis between WLI and NBI, the pooled sensitivity and specificity was estimated to be 82% [79.20; 91.80, 95% CI, p < 0.01, I^2 = 98.0% [97.5%; 98.4%] ] and 92.16 [87.85; 95.02, 95% CI, p < 0.01, I^2 = 97.3% [96.5%; 97.9%] utilising the random effects model. The subgroup analysis based on the type of imaging modality indicated that narrow-band imaging ( 94% and 98%) showed higher sensitivity and specificity in comparison to white light imaging models (74% and 92%). Conclusions: The evidence from this meta-analysis clearly demonstrates the robust diagnostic performance of deep learning models. Region-based CNN models offered superior diagnostic performance compared to conventional CNN models. These findings highlight the potential of region-based CNN, suggesting further exploration. The evidence also indicates the superiority of narrow band imaging as a diagnostic modality of gastric cancers compared to regular white light imaging.