Diagnostic Accuracy for Gastric Cancer, Adenoma, and Intestinal Metaplasia With vs Without AI Assistance: an observer-based, reader-blinded, randomized case-order exploratory validation study

Background Artificial intelligence (AI) can accurately classify gastric lesions, but its clinician-level impact in real-world practice remains uncertain. We compared endoscopists’ diagnostic performance with vs without AI assistance using both still-image (M1) and video (M2) datasets. Methods We analyzed 1,570 cases (226 cancer, 282 adenoma, 413 non-neoplastic lesions [NNL], 297 intestinal metaplasia [IM], and 352 gastritis/normal). One representative still image per case was extracted for M1; edited five-second video clips formed M2. Six in-training endoscopists (< 3 years’ experience) independently read M1 and M2 with and without AI after a one-week washout. Results As a stand-alone model, AI achieved 91.31% (M1) and 92.51% (M2) accuracy for focal lesions (sensitivities 91.02% and 91.91%; specificities 95.50% and 96.12%). For IM, accuracy was 91.83% (M1) and 92.45% (M2). With AI assistance, overall reader accuracy increased from 74.80% to 86.84% in M1 (AUC 0.742 to 0.860) and likewise from 74.80% to 86.84% in M2 (AUC 0.796 to 0.900); all p  < 0.05. By subtype (videos, M2), accuracy improved 80.01% to 89.85% for cancer (+ 9.84%), 44.86% to 63.77% for adenoma (+ 18.91%), 72.34% to 84.26% for NNL (+ 11.92%), and 66.36% to 89.30% for IM (+ 22.94%). Still-image results showed similar gains (e.g., adenoma 44.86% to 64.95%, IM 66.36% to 86.48%, both p < 0.05). Conclusions AI assistance significantly enhances endoscopists’ diagnostic accuracy across lesion types and modalities, with the largest benefits for adenoma and IM—conditions prone to clinician-level variability. These findings support integrating AI into routine upper endoscopy to improve diagnostic reliability and earlier recognition of clinically significant lesions.