Ampullary neoplasms are often challenging to detect during forward-viewing esophagogastroduodenoscopy. This study aimed to evaluate the impact of artificial intelligence (AI) assistance on the accuracy of endoscopists in identifying ampullary lesions. The stand-alone AI model achieved an accuracy of 84.2%, with AUCs ranging from 83.2% to 90.9% across binary classification tasks. With AI assistance, mean diagnostic accuracies significantly improved for endoscopy fellows (63.5%-75.8%, P < 0.001) and gastrointestinal endoscopists (68.4%-76.5%, P = 0.012), while no significant difference was observed for pancreatobiliary endoscopists (70.8%-73.1%, P = 1.000). In binary tasks, AI assistance significantly enhanced specificity for two fellows in distinguishing normal from abnormal ampullae (P = 0.023 and P = 0.041) and for another fellow in differentiating malignancy (67.4%-89.1%, P = 0.004). These improvements in accuracy and specificity, particularly among less-experienced endoscopists, were achieved without compromising sensitivity for the detection of abnormal lesions. AI assistance enhanced diagnostic accuracy and specificity for ampullary lesions, particularly among non-pancreatobiliary endoscopists, without compromising sensitivity. These findings demonstrate the retrospective proof-of-concept potential of AI assistance, highlighting the need for subsequent prospective and video-based validation to confirm its clinical utility in the evaluation of ampullary lesions.
Seo et al. (Wed,) studied this question.