Artificial intelligence is demonstrating high diagnostic performance in identifying early gastric cancer (EGC) from white light endoscopy images, according to a recent systematic review and meta-analysis. Deep learning algorithms have been found to match the diagnostic accuracy of expert endoscopists.
Early gastric cancer, defined as adenocarcinoma that infiltrates the stomach, carries a favorable prognosis and a 5-year survival rate of approximately 95% when detected early. Upper gastrointestinal endoscopy is the established gold standard for diagnosing EGC, with white light endoscopy being the preferred technique due to its accessibility and ease of use. While screening endoscopy has significantly reduced mortality, the intricate nature of EGC lesions makes detection highly reliant on endoscopic expertise.
A comprehensive analysis of 15 studies, encompassing over 37,000 white light endoscopy images, revealed impressive results. In internal validation, deep learning algorithms correctly identified 91% of patients with EGC and accurately ruled out 93% of those without. For external validation, DL models diagnosed 82% of EGC cases and ruled out 83% of non-EGC cases. Crucially, the study found no significant difference in diagnostic sensitivity and specificity between the AI models and expert endoscopists.
While these findings are promising, researchers caution that the studies utilized retrospective datasets, potentially leading to a "best-case" scenario. The complexity of real-time endoscopy, with factors like motion blur and illumination variations, is not fully replicated in static image analysis. Nevertheless, deep learning algorithms hold substantial potential as clinical decision-support tools to enhance routine practice in early gastric cancer diagnosis.