Deep learning model for diagnosing gastric mucosal lesions using endoscopic images: development, validation, and method comparison

Background and Aims: Endoscopic differential diagnoses of gastric mucosal lesions (benign gastric ulcer, early gastric cancer [EGC], and advanced gastric cancer) remain challenging. We aimed to develop and validate convolutional neural network–based artificial intelligence (AI) models: lesion detection, differential diagnosis (AI-DDx), and invasion depth (AI-ID; pT1a vs pT1b among EGC) models. Methods: This study included 1366 consecutive patients with gastric mucosal lesions from 2 referral centers in Korea. One representative endoscopic image from each patient was used. Histologic diagnoses were set as the criterion standard. Performance of the AI-DDx (training/internal/external validation set, 1009/112/245) and AI-ID (training/internal/external validation set, 620/68/155) was compared with visual diagnoses by independent endoscopists (stratified by novice [<1 year of experience], intermediate [2-3 years of experience], and expert [>5 years of experience]) and EUS results, respectively. Results: The AI-DDx showed good diagnostic performance for both internal (area under the receiver operating characteristic curve [AUROC] =.86) and external validation (AUROC =.86). The performance of the AI-DDx was better than that of novice (AUROC =.82, P =.01) and intermediate endoscopists (AUROC =.84, P =.02) but was comparable with experts (AUROC =.89, P =.12) in the external validation set. The AI-ID showed a fair performance in both internal (AUROC =.78) and external validation sets (AUROC =.73), which were significantly better than EUS results performed by experts (internal validation, AUROC =.62; external validation, AUROC =.56; both P <.001). Conclusions: The AI-DDx was comparable with experts and outperformed novice and intermediate endoscopists for the differential diagnosis of gastric mucosal lesions. The AI-ID performed better than EUS for evaluation of invasion depth.