Detection of large flat colorectal lesions by artificial intelligence: a persistent weakness and blind spot

Abstract

Computer-aided detection (CADe) has increased adenoma detection in randomised trials. However, unlike other detection adjuncts, CADe is lesion specific, that is, it is trained on a specific set of lesions. If the training does not include sufficient examples of precancerous lesion subsets, CADe may not perform adequately for lesions in that subset. In a prospective assessment of a second-generation CADe programme in 165 colonoscopies, we identified 26 flat lesions ≥10 mm in 17 patients. The endoscopist identified 22 of 26 lesions before the CADe programme. In 13 lesions, the CADe either generated no detection signal or only a signal over part of the lesion after colonoscope position or luminal inflation adjustment. Thus, the second-generation CADe algorithm, like the first generation, frequently fails to effectively detect large flat colorectal lesions, which are likely very important lesions that a CADe programme should identify. The first CADe programme to be launched commercially in the USA was GI Genius (Medtronic, Minneapolis, Minnesota, USA). We showed that this CADe programme failed to generate a detection signal disproportionately for large flat lesions, particularly large serrated lesions,1 though CADe appears to overall improve detection of sessile serrated lesions.2 A new Medtronic CADe programme termed ‘ColonPRO’ offers new artificial intelligence (AI)-based features with potential value including automatic measurement of the Boston Bowel Preparation Score and various procedure-related times. In an initial assessment of this new programme, we recorded the frequency of an AI-based signal for large (≥10 mm) flat lesions. We conducted the evaluation as a quality assessment. Permission to report the findings was granted by the Indiana University Institutional Review Board on 15 October 2024. In a single centre, after exclusion of patients with familial polyposis and inflammatory bowel disease, ColonPRO was activated throughout 165 consecutive colonoscopies. In the 165 study patients, 18 lesions referred …