Prevention of colorectal adenocarcinoma mainly relies on the detection and resection of precancerous lesions (adenomas and serrated lesions) during colonoscopy. The implementation of artificial intelligence (AI) solutions in digestive endoscopy aims to improve detection performance. However, in the literature, detection rates depend on multiple factors related to the patient, the operator, and the procedural context. Moreover, while AI seems to increase the number of resected lesions, its impact on the detection of clinically significant lesions remains debated. The primary objective of this study was to evaluate the impact of AI on the detection of polyps ≥10 mm.

We conducted a prospective, observational, single-center study in a university hospital between January and December 2024. Colonoscopy data were collected by operators using dedicated questionnaires. Inclusion of examinations and the choice to use AI were left to the operator’s discretion. Patient and procedural characteristics as well as quality indicators were recorded. Operator perception of AI was also assessed. A propensity score matching analysis was performed on age, sex, and operator experience to evaluate factors associated with resection of polyps ≥10 mm, after excluding known lesions and colonoscopies performed for inflammatory bowel disease. The study protocol was approved by the local ethics committee.

A total of 547 colonoscopies (median age 63 years; 57% male) were analyzed, including 290 with AI and 257 without AI. Indications differed significantly between groups, with more follow-up colonoscopies (32% vs. 17%) and FIT-positive indications (8% vs. 4%), and fewer known lesions (14% vs. 4%) in the AI group compared to the non-AI group (p<0.01). Colonoscopies with AI were more often performed by experienced endoscopists, in the morning, and with patients in the left lateral position. The total number of resected polyps and detection of polyps ≥10 mm did not differ between AI and non-AI groups. However, total procedure time (20 vs. 25 min) and withdrawal time (10 vs. 13 min, p<0.05) were significantly shorter with AI. After propensity score matching, the proportion of colonoscopies resecting at least one polyp ≥10 mm was not significantly different between AI and non-AI groups (20.7% vs. 20.6%).

In multivariable analysis, factors associated with detection of polyps ≥10 mm were operator status (assistant or fellow vs. resident; OR 0.04; p=0.004), absence of music in the endoscopy room (OR 0.2; p=0.013), screening colonoscopy vs. symptomatic colonoscopy (OR 0.08; p<0.001), and longer withdrawal time (OR 14.7; p<0.001). AI use had no impact on detection of polyps ≥10 mm.

Operator perception from the 290 AI-assisted procedures showed that 88.3% of endoscopists considered AI neither saved nor wasted time. System interruptions were rare (2.1%). AI was reported to detect a polyp missed by the operator in 5.2% of cases, but failed to detect a polyp seen by the operator in 7.6%.

In this prospective real-life study, AI did not improve detection of clinically significant lesions. Detection performance remained mainly determined by patient- and operator-related factors. In routine practice, numerous confounding factors exist, which differ from the highly controlled conditions of randomized trials.