Authors (including presenting author) :
Thomas KL Lui(1), Cynthia Hui(1), Vivien WM Tsui(1), Michael KS Cheung(1), Michael KL Ko(1), CC Fu(2), CK Yeung(3), SY Wong(1), Wai K Leung(1)
Affiliation :
(1)Department of Medicine, Queen Mary Hospital, University of Hong Kong(2)Department of Surgery, Queen Mary Hospital, University of Hong Kong,(3) NISI (HK) Limited
Introduction :
A recent meta-analysis showed that up to 26% of adenoma were missed during colonoscopy. These missed adenomas are the main cause of interval colorectal cancer after colonoscopy
Objectives :
We investigated whether the use of artificial intelligence (AI) assisted real-time detection could reduce missed colonic lesions during colonoscopy.
Methodology :
A deep learning AI model for the real time detection of colonic neoplasm with high accuracy had been developed. The AI model was first validated in the videos of tandem colonoscopy of the proximal colon (from caecum to splenic flexure) collected in our on-going study. The gold standard was all lesions detected in the proximal colon by the endoscopists in the tandem examinations with histological confirmation. We then validated the real time AI model in prospective colonoscopy examinations in patients undergoing colonoscopy for various indications. The colonoscopies were performed by both experienced endoscopists and junior endoscopists. The AI findings were displayed in a separate monitor and observed by an independent investigator, while the endoscopists was blinded to the real-time AI findings. The colon was divided into three segments (right-side, transverse and left-side colon) and segmental unblinding of the AI result would be provided after complete examination of each segment. If a lesion was detected by AI but not by the endoscopists, that segment would be re-examined. If no additional lesion was detected by AI, the endoscopists would proceed to examination of next segment. All polyps detected were removed for histological examination.
Result & Outcome :
In the retrospective review of 65 tandem colonoscopy videos of the proximal colon, there were 32 polyps (24 adenomas) missed in the first examination and were detected on second examination only. AI could detect 78.1% of these missed polyps (79.1% of missed adenoma) in the video of the first-pass examination and 100% of lesions were detected by AI on tandem examination. In the prospective study of 52 colonoscopy examinations (mean patient’s age 66.2 years; 51.9% male), the overall adenoma detection rate was 67.3% and a total of 148 polyps (94 adenomas) were detected with AI assistance. There were 46 (31.1% of total) missed polyps or 18 (19.1% of total adenoma detected) missed adenomas detected by AI. AI could detect at least one missed adenoma in 26.9% of the cases. With the AI assistance, the total number of polyps and adenomas detected were increased by 32.1% and 23.6%, respectively. Six (14.6%) suspected missed lesions detected by the AI could not be identified on reexamination. Multivariable analysis showed that the presence of missed adenoma(s) detected by AI was more likely with less experienced endoscopists (adjusted OR 1.30 95%CI 1.05-1.62, p=0.02) and increase in the total number of colonic polyps (adjusted OR 1.05, 95%CI 1.02-1.09, p< 0.001).
AI assisted real-time detection could potentially detect 79% of the missed proximal adenoma which were detected in tandem examinations. It could also reduce the numbers of missed adenoma in a single examination in up to 26% of cases. Missed adenoma detected by AI seems more likely happened when there were multiple colonic polyps and performed by less experienced endoscopists.
