Does AI-assisted colonoscopy improve adenoma detection in screening colonoscopy? A large-scale multi-centre randomised controlled study

INTRODUCTION: Recent studies suggested that artificial intelligence (AI)‐assisted colonoscopy is a promising technology that may improve polyp detection and characterization in colonoscopy. However, data from large scale multi‐centre randomised controlled trial assessing the efficacy of real time AI assistance in the detection of polyps/adenoma in a screening population is lacking. AIMS & METHODS: This is a randomised controlled trial comparing the efficacy of AI‐assisted colonoscopy in adenoma detection to that of standard colonoscopy in a screening population in 6 referral centres in Hong Kong, Jilin, Inner Mongolia, Xiamen and Beijing. Asymptomatic subjects aged 45 to 75 years undergoing CRC screening by direct screening colonoscopy or fecal immunochemical test (FIT) based screening program regardless of their FIT results were invited for the study. Eligible subjects were randomized to AI group and standard group (control) in a 1:1 ratio. In AI group, a convolutional neural network‐based, polyp detection system (Eagle‐Eye, Xiamen InnoVision) with real‐time visual notice on the same monitor of the standard high‐definition endoscopy system was used during intubation and withdrawal. Polypectomy are allowed for polyps <10mm during insertion in both arms. Both junior (< 5000 colonoscopies) and senior endoscopists (≥ 5000 colonoscopies) performed the study procedures in each centre. The primary outcome was overall ADR. Secondary outcomes were mean number of adenomas per colonoscopy (APC), difference in ADR according to the experience of endoscopists, and colonoscopy withdrawal time. RESULTS: From November 2019 to March 2021, 1799 subjects were randomised (902 in AI group and 897 in standard group). Baseline characteristics were similar between the two groups. Median Boston Bowel Preparation Scale (BBPS) score (IQR) was 6 (6‐8) in both groups. The overall ADR (40.8% vs 35.1%, p=0.013) and the mean number of APC (0.62 vs 0.51, p=0.011) of AI group were significantly higher than those of the standard group. The right colon ADR (15.7% vs 11.3%, p=0.005) and the left colon ADR (22% vs 18.1%, p=0.039) were both significantly higher in the AI group. However, there was no difference in the transverse colon ADR (11.5% vs 13.4%, 0.235). There were trends toward higher overall ADR in both senior (41.6% vs 36.1%, p=0.088) and junior (40% vs 34.2%, p=0.077) endoscopists with AI‐assistance but without reaching statistical significance. The median withdrawal times (minutes) [IQR] (8.2 [6.7‐12.2] vs 7.9 [6.5‐11.2], p=0.125) were similar in both groups while the median intubation time (minutes) [IQR] was slightly longer in the AI group (4.5 [3.3‐6.1] vs 4.1 [3.2‐5.5], p<0.001). CONCLUSION: In this large‐scale multi‐centre randomised controlled study involving both junior and senior endoscopists, real time AI‐assisted colonoscopy improved overall, right and left colon ADR in screening colonoscopy.