What an AI polyp alert changes during a colonoscopy

An AI alert during a colonoscopy is a request for a closer look. A system can place a coloured box around a suspicious patch in the live camera image, drawing the clinician’s attention to something worth examining. The clinician still decides what the finding means and whether to remove tissue. That division of labour is central to the original FDA description of GI Genius.[1]

The underlying examination remains familiar. A flexible tube carries a camera through the large bowel, following preparation to empty it. A clinician can remove growths called polyps or take samples during the procedure. Testing those findings can then inform treatment or a later check-up. AI enters one moment in that longer chain: the search for an abnormality that deserves attention.[2]

The alert and the diagnosis

GI Genius’s original design uses a convolutional neural network to analyse video and add green boxes over potential lesions. The camera image remains available to the endoscopist. Recognising a visual pattern and deciding its clinical significance are separate jobs; the device’s instructions preserve that distinction. A box is an invitation to inspect, rather than a pathology result.[1]

The distinction also appears in NICE’s terminology. Computer-aided detection, or CADe, helps locate a polyp. Computer-aided characterisation, or CADx, estimates what kind it is. In its March 2026 guidance, NICE gave these functions different recommendations because the evidence supporting them differs.[8]

What counts as finding more?

The adenoma detection rate, usually shortened to ADR, counts examinations in which at least one adenoma is found. Some adenomas can develop into cancer, which makes detecting and removing them valuable. Another measure, adenomas per colonoscopy, counts how many are found in total, divided by the number of examinations. ADR counts affected examinations; the average captures additional findings in the same patient.[4]

Consider an invented set of ten examinations. Four reveal at least one adenoma, and six adenomas are found altogether. ADR is 40%, while adenomas per examination is 0.6. Now suppose a better search finds three extra adenomas, all in those same four people. ADR remains 40%; the average becomes 0.9. The extra discoveries have disappeared from the first measure because none occurred in a previously negative examination. This is why a single headline percentage can miss part of a detection tool’s contribution.

Three studies, three useful perspectives

A 2022 trial led by Michael Wallace used a particularly revealing design: 230 patients underwent two examinations on the same day, with the order randomised between AI-assisted and standard colonoscopy. An adenoma missed initially could be discovered on the second pass. With AI first, 38 of 246 adenomas were missed; with standard examination first, 80 of 247 were missed. The denominators are all adenomas found across the two passes in each group. These percentages measure lesions overlooked on the first examination, rather than the proportion of patients who would develop cancer.[3]

The Danish study offers a view of routine practice. AI and control periods alternated in two-week blocks at four centres. About 40% of patients came through screening after a positive stool blood test. ADR rose alongside adenomas per colonoscopy, from 1.11 to 1.35. One missing histology result left 399 AI patients for the adenoma calculation, compared with 395 controls. The calendar-based allocation and enriched screening population help explain what the result represents.[4]

EndoMind tells a different story. Its randomised private-practice study, published in 2026 using 2021–22 data, found no significant ADR improvement. Ten highly experienced examiners performed the procedures. Median withdrawal time was similar: 7.5 minutes with AI and 7.6 without. The study therefore does not describe a situation in which simply spending much longer looking produced the AI result.[5]

Its limitations also matter. One centre was excluded because of documentation and setup problems, reducing the available sample. The control group detected more adenomas than the original sample-size calculation had expected. The result is a useful account of this particular system and clinical setting; it does not settle how much every detection aid adds in every endoscopy unit.[5]

Does the clinician learn from the assistant?

That question becomes more interesting after months of use. A 2025 observational study at four Polish centres compared colonoscopies performed without AI before and after the technology entered practice. ADR fell from 28.4% among 795 examinations to 22.4% among 648. The association raised a concern about dependence on assistance. Because the comparison was observational, it did not establish that clinicians had lost a lasting ability to recognise adenomas.[6]

A Norwegian study published in 2026 followed 5,013 colonoscopies by 13 endoscopists across periods before AI, during use and after removal. Among the less experienced operators, detection of at least one polyp measuring 5 mm or more rose from 31.9% to 39.5% while AI was present. After its removal, the adjusted change from baseline was not significant. This endpoint was polyp detection by size, not histologically confirmed ADR.[7]

The Norwegian authors identify a useful difference between the experiments. Poland measured unaided procedures during a period when clinicians also performed AI-assisted ones. Norway examined performance after the assistance period ended. Those arrangements ask different questions about attention and experience. Together, they make the case for measuring how the clinician and tool perform over time, alongside the immediate effect of switching an overlay on.[7]

From a detected polyp to better care

NICE’s March 2026 decision reflects both the promise and the remaining work. Six named detection technologies can be used in the NHS for four years while evidence is generated, subject to specified conditions, for people without diagnosed inflammatory bowel disease or Lynch syndrome. Those groups need further research, as does computer-aided characterisation. Evidence generation should follow the size and type of additional findings, subsequent management and cancers diagnosed after colonoscopy. Detecting more small adenomas is useful information, but does not itself provide a measured reduction in later cancer.[8]

For someone having the examination, the meaningful result is what follows the search. The NHS explains that findings are tested and may lead to treatment or a future colonoscopy. Useful questions therefore extend beyond whether AI was present: what was found, what was removed or sampled, and what happens next? The alert earns its place by helping the clinical team make those decisions from a more complete examination.[2]

Report an error or suggest a correction