AI did not raise adenoma detection in Lynch syndrome colonoscopy, a trial found

Computer-aided detection (CADe) is now one of the better-evidenced uses of AI in medicine: software watches the colonoscopy video in real time and boxes anything that looks like a polyp, and in ordinary average-risk screening it reliably finds more adenomas than the endoscopist alone [s2]. The open question is whether that benefit carries over to the people who need it most. A randomised trial in patients with Lynch syndrome — an inherited condition that sharply raises colorectal cancer risk and commits carriers to lifelong surveillance — set out to answer it, and found that it did not [s1].

What CADLY2 tested

CADLY2 was an international, multicentre, open-label, randomised superiority trial run at nine specialised hereditary-cancer surveillance centres in Belgium, Germany, the Netherlands and Spain [s1]. Adults with genetically confirmed Lynch syndrome who were due for surveillance colonoscopy were randomly assigned 1:1 to high-definition white-light colonoscopy alone or to the same with assistance from CAD EYE, a commercial system made by Fujifilm, which flagged lesions during withdrawal and then offered an optical "neoplastic or not" call once a lesion was found [s1]. The trial's funding came from third-party research support of the National Center for Hereditary Tumor Syndromes at University Hospital Bonn, not from the device-maker [s1].

Between 9 May 2023 and 30 October 2025, 757 patients were randomised — 377 to standard colonoscopy and 380 to AI-assisted colonoscopy — with 733 in the full analysis set (369 standard, 364 AI) [s1]. The median age was 49 years (IQR 38–59) in the standard group and 50 in the AI group [s1]. The primary outcome was the adenoma detection rate: the proportion of patients in whom at least one adenoma was found and confirmed by pathology [s1].

The result

The adenoma detection rate was 30.9% (114 of 369 patients) with standard colonoscopy and 33.8% (123 of 364 patients) with AI assistance — an odds ratio of 1.14 (95% CI, 0.83–1.57; p=0.41) [s1]. The numerical difference points the AI's way, but the confidence interval spans no effect, and the trial did not reach the absolute improvement its sample size had been built to detect [s1].

The optical-diagnosis half fared no better against the standard it needed to beat. For telling neoplastic from non-neoplastic lesions, the AI's sensitivity was 85.9% (95% CI, 82.0–89.1) and specificity 91.4% (89.4–93.0) [s1] — respectable numbers in the abstract, but the authors concluded the system did not clearly improve on expert optical diagnosis in these specialist centres [s1]. Three adverse events occurred in the AI group — two mild post-polypectomy bleeds and one pulmonary embolism or deep-vein thrombosis judged unrelated to the procedure — and none in the standard group [s1].

The trial was built to find a benefit, not to rule one out. It was designed as a superiority trial and powered around an absolute improvement in adenoma detection that its sample size assumed the AI would deliver [s1]. When the observed gap between 30.9% and 33.8% came in smaller than the calculation presumed, the study was left without the statistical room to call that difference real — which is why a result that looks directionally favourable reads, correctly, as a negative trial [s1]. The randomisation itself was careful: allocation was concealed through a central web system and stratified by centre, sex, previous colorectal cancer, the underlying genetic variant and the interval since the last colonoscopy [s1].

Why the screening benefit may not transfer

That CADe helps in average-risk screening is not in doubt. A 2026 network meta-analysis of 21 randomised trials and 19,006 participants found several systems significantly out-detected conventional colonoscopy — ENDO-AID at a rate ratio of 1.27 (95% CrI, 1.13–1.41), Eagle-Eye at 1.19 (1.02–1.38) and DEEP² at 1.37 (1.07–1.76) for adenoma detection [s2]. So why would a strong tool fall flat in Lynch syndrome?

The likeliest answer is headroom. CADe prevents the misses an ordinary endoscopist makes under time pressure; in a dedicated hereditary-cancer centre, the endoscopists are already looking slowly and expertly for exactly the lesions Lynch syndrome produces, which are often flat and subtle rather than the bulky polyps CADe is trained to catch. When the human baseline is already high, an accurate second reader has little left to add — the same pattern this site has seen when an accurate fracture-spotting AI failed to change what happened to patients. A device that improves care in one setting is not thereby proven to improve it in another.

What to watch

Whether CADe is retested in Lynch syndrome at non-specialist centres, where the surveillance actually happens for many carriers and where the human baseline is lower — the setting most likely to reveal a benefit if one exists. CADLY2 does not show the technology is useless here; it shows that in expert hands, on this specific population, the gain assumed from screening data did not materialise [s1]. The lesson is the recurring one of clinical AI: the benefit is a property of the setting, not of the software.

This article is informational and is not medical advice.

Sources

• [s1] Hüneburg R, van Bokhorst QNE, Pellisé M, et al. "Artificial intelligence-assisted detection and optical differentiation of colorectal lesions in Lynch syndrome surveillance (CADLY2): a multicentre, open-label, randomised controlled superiority trial." The Lancet Gastroenterology & Hepatology, 11(10):875–886, published online 16 July 2026. https://doi.org/10.1016/S2468-1253(26)00163-9
• [s2] "Artificial intelligence-based computer-aided detection systems for adenomas during colonoscopy: a network meta-analysis." Frontiers in Artificial Intelligence, published online 9 September 2026. https://doi.org/10.3389/frai.2026.1880524