Abstract
BACKGROUND AND AIMS: Multiple artificial intelligence (AI) systems have been developed to assist with endoscopic diagnosis. We established the first real-time AI lesion-labeling system to assist in delineating lesion margin during endoscopic submucosal dissection (ESD). We aimed to further validate the efficacy of this system in improving histological complete resection rate especially for beginners in low-volume centers. METHODS: We performed this prospective cohort study in two endoscopy centers in Shanghai, China from January 2021 to December 2022. Eligible patients in a low-volume center equipped with real-time AI lesion labeling system were recruited consecutively to the AI group and underwent ESD performed mainly by beginners, while participants in a high-volume center underwent conventional ESD performed by experienced endoscopists. The primary outcome was complete lateral resection rate according to pathological examination of the specimen resected. Secondary outcomes were en -bloc resection rate, procedure duration, specimen diameter and complication rate. RESULTS: 174 patients (200 lesions) were recruited into the AI-assisted ESD group. With the use of lesion-margin labeling system, 90.0% (180/200) of ESD cases achieved negative lateral margins. The en bloc resection rate was 98.5% (197/200), and the histological complete resection rate was 87.5% (175/200). 181 patients (202 lesions) received conventional ESD in a high-volume center. There was no significant difference between AI-assisted and conventional ESD group regarding the rate of complete lateral resection (90.0% [180/200] vs 92.1% [186/202], P = 0.465). Total procedure duration (min) was significantly longer in AI-assisted group (82 [54,106]) compared with conventional group (49 [30,63], P < 0.001). CONCLUSION: AI lesion-labeling system showed reliable safety and efficacy in assisting to identify the margin of lesions before ESD, and potential to improve complete lateral resection rate of ESD for esophageal lesion in low-volume endoscopy centers.