A computer that grades each colonoscopy raised how often endoscopists found adenomas
A Danish stepped-wedge trial gave endoscopists automated feedback on their technique after every procedure. Adenoma detection rose from 43.4% to 48.6% — a different tool from real-time polyp AI.
| Group | Value (%) |
|---|---|
| Adenoma detection — feedback | 48.6 |
| Adenoma detection — control | 43.4 |
| Polyp detection — feedback | 54.2 |
| Polyp detection — control | 50.7 |
| Adenocarcinoma detection — feedback | 8.9 |
| Adenocarcinoma detection — control | 6.5 |
Most artificial intelligence in colonoscopy has aimed at the moment of the procedure itself — software that watches the video feed and flags a polyp as the scope passes it. A trial published in The Lancet Digital Health on 2 September tests a different idea: not an AI that spots lesions in real time, but a system that grades the endoscopist's technique after each colonoscopy and feeds the scores back to them [s1].
Why detection rate matters
The outcome the trial targets is the adenoma detection rate (ADR) — the proportion of colonoscopies in which at least one precancerous adenoma is found. ADR is the most established measure of colonoscopy quality, and it matters because it is tied to outcomes: a landmark cohort found that each one-percentage-point increase in a physician's ADR was associated with roughly a 3% lower risk of a subsequent ("interval") colorectal cancer [s2]. Anything that reliably raises ADR is therefore of direct interest for cancer prevention.
What the trial did
The study was a cluster-randomised, controlled trial with stepped-wedge randomisation across three university hospitals in Denmark, enrolling patients aged 50–74 with a positive faecal immunochemical test [s1]. The three hospital sites were the units of randomisation, each crossing over from a control period to the intervention period at a scheduled point [s1].
During the control period, endoscopists performed standard colonoscopies [s1]. During the intervention period, they received computer-aided quality feedback after each procedure, scoring metrics defined by the "Copenhagen Colonoscopy Concept" — a retraction score covering how the scope was withdrawn, plus cecal intubation rate, intubation time, number of polyps detected, withdrawal time and distance to cecum [s1]. The idea is that seeing objective marks on their own technique nudges endoscopists to improve it.
Between 23 February 2022 and 19 April 2024, the final analysis included 1,109 patients in the intervention group and 1,001 in the control group; 44.3% were female and 55.7% male [s1]. Seventeen endoscopists took part and acted as their own controls [s1].
The results
The feedback period had a higher adenoma detection rate than the control period — 48.6% versus 43.4% (odds ratio 1.24, 95% CI 1.02 to 1.52; p=0.032) [s1]. The polyp detection rate rose from 50.7% to 54.2% (OR 1.22, 95% CI 1.00 to 1.49; p=0.048), and the adenocarcinoma detection rate from 6.5% to 8.9% (OR 1.51, 95% CI 1.06 to 2.15; p=0.027) [s1]. Endoscopists also spent longer withdrawing the scope — 960 versus 915 seconds (incidence rate ratio 1.11, 95% CI 1.06 to 1.17; p=0.003) — consistent with more careful inspection [s1].
Not every measure moved. The mean number of adenomas detected per colonoscopy was not significantly different (1.00 versus 0.85; IRR 1.15, 95% CI 0.99 to 1.33; p=0.069), nor were the per-colonoscopy counts of tubular adenomas or polyps [s1]. So the intervention raised the proportion of colonoscopies that found something, more clearly than it raised the number of lesions found each time.
What to make of it
This is a genuine randomised trial with a clinically meaningful endpoint, and its result is coherent: feedback lengthened withdrawal time and lifted detection rates, exactly the causal chain you would hope for. The gain in adenocarcinoma detection is notable, though the wide confidence interval reflects how few cancers there are to count.
The limits are real. Endoscopists could not be blinded to whether they were receiving feedback, and simply knowing you are being measured can change behaviour — a Hawthorne effect that is impossible to separate from the feedback content itself. With three sites and 17 endoscopists, the trial is modest, and it was conducted in a faecal-test-based screening programme; the authors note the effect should be explored in centres that do not use such tests. Most importantly, the trial measured detection, not the outcome that matters to patients — whether fewer interval cancers actually occur — which the authors flag as the necessary next study [s1].
Read against the wider field, the interest of this trial is its target. Rather than assisting the eye in the moment, it works on the operator's habits over time. Whether post-hoc quality feedback and real-time detection AI are competing or complementary is a question the evidence has not yet answered.
Sources
- Effect of a computer-aided quality feedback system on colonoscopists' adenoma detection rate in Denmark — The Lancet Digital Health, 2026-09-02
- Adenoma Detection Rate and Risk of Colorectal Cancer and Death — New England Journal of Medicine, 2014
Sources
- Effect of a computer-aided quality feedback system on colonoscopists' adenoma detection rate in Denmark: a multicentre, stepped-wedge, cluster-randomised, controlled trial — The Lancet Digital Health , September 2, 2026
- Adenoma Detection Rate and Risk of Colorectal Cancer and Death — New England Journal of Medicine , April 3, 2014
More on
Polyp-detection AI moved to the cloud. The lag was 59 milliseconds; the detection tripled.
The EAGLE trial ran colonoscopy AI off-site over a network and aimed it at the lesions that matter. It reports a threefold gain in serrated lesions — in a field whose main US guideline recommends nothing.
Screening guidelines disagree because the same trials support more than one answer
US guidance recommends mammography from 40. A Canadian review of the same randomised evidence put the benefit at 0.27 fewer breast cancer deaths per 1,000 women aged 40-49 over ten years.
A 278,000-person screening trial found more early cancers — and more early harms
SCREESCO randomised Swedish 60-year-olds to colonoscopy, stool testing, or nothing. After a median 4.8 years, overall cancer incidence barely moved; the stage distribution did.
Processed meat is a Group 1 carcinogen. That is a statement about evidence, not risk.
IARC put bacon in the same category as tobacco in 2015, which says how sure scientists are, not how dangerous it is. The attributable-death figures are three orders of magnitude apart.