Across 314,872 colonoscopies, the detection rate for precancerous lesions ranged from 7.4% to 52.5% depending on the gastroenterologist. Equipment matters, but the experience and the results of the operator can radically change what is found.
If an examination has left you unconvinced, or if you would like its quality checked, Noria Health Hub can arrange a full check-up or a second medical opinion in Brussels, with a report within 24 h, coordinate your care where needed across more than 40 medical pathways in 48 h, then, if your medical condition allows, arrange surgery within 7 days.
A 7-fold spread between operators
In 2014 the New England Journal of Medicine published an analysis of 314,872 colonoscopies performed by 136 gastroenterologists. The aim was straightforward, to measure the adenoma detection rate, that is the proportion of examinations in which at least one precancerous lesion is found and removed.
The result is one of the most striking in modern medicine. Depending on the operator, that rate ranges from 7.4% to 52.5%. A 7-fold spread, for the same examination, within the same care organisation, with the same equipment.
The gap does not stay on paper. The same study followed what happened to the patients. Among those examined by an operator in the top quintile, compared with the bottom quintile, observed interval cancer was close to 2 times less frequent, and death from that cancer more than 2 times less frequent. The study is observational, it measures an association and not an effect. Every additional point of detection rate comes with a 3% fall in cancer risk.
A known limitation: about one adenoma in four is missed
A meta-analysis published in Gastroenterology in 2019 pooled 43 publications and more than 15,000 colonoscopies performed twice in a row in the same patients, which makes it possible to count exactly what the first examination let through. The adenoma miss rate is 26%. One in four. For advanced adenomas, 9%. For sessile serrated polyps, 27%.
This miss rate is a known limitation of colonoscopy. What separates one centre from another is therefore mainly the means used to reduce that risk: the experience of the operator, the quality of the preparation, the time spent on the examination and the assistance tools available.
The same variability exists in imaging
Radiology tells the same story. An analysis of 205 radiologists and more than a million screening mammograms, published in Radiology in 2009, found a median sensitivity of 83.8%, with an interquartile range from 74.5 to 92.3%. The single factor associated with better detection was sub-specialisation in breast imaging.
A second analysis, published in Radiology in 2011 across 120 radiologists and nearly 784,000 mammograms, shows that false positive rates are significantly higher among lower-volume readers. Concentrated experience cannot be substituted.
What artificial intelligence improves, and what it has not shown
One question remains: does computer assistance narrow these gaps between operators? The trials give different results depending on the setting, so the positive and the negative studies have to be read together.
| Trial | Scale | Result | Publication |
|---|---|---|---|
| COLO-DETECT | 12 UK hospitals, 2,032 patients | Detection from 48.4% to 56.6% | Lancet Gastroenterology and Hepatology, 2024 |
| Tandem trial | 8 centres | Missed lesions 15.5% against 32.4%, 2 times fewer | Gastroenterology, 2022 |
| TIMELY | 17 European centres | No benefit | Lancet Gastroenterology and Hepatology, 2024 |
| Swedish trial | 78% of procedures by experienced endoscopists | No overall gain | 2024 |
| US pragmatic trial | Routine practice | No benefit | Gastroenterology, 2023 |
This table includes the positive trials as well as the negative ones, so that the full set of available results is visible.
Taken together, these trials mainly suggest that computer assistance can help pick up lesions that would otherwise have been missed, particularly where the baseline performance is not optimal. It does not replace the operator, and its benefit is less obvious where detection is already very good.
What the studies do not allow us to claim
To date, no randomised trial has shown that an artificial intelligence diagnostic tool reduces cancer mortality, in any organ. What is measured is lesions detected, not lives saved. We use these tools because missing less has value in itself. We will not claim that they cure you.
In practical terms, if this concerns you
You set out your problem, your symptoms or your question. Reports, test results and images already available can be studied before you arrive, so that the pathway is prepared. The aim is not to run every possible examination, but to select those that can genuinely help confirm a hypothesis, rule out a risk or guide a decision. The findings are then brought together and the next steps are organised. Follow-up is coordinated from there.
Need a medical opinion quickly? For a concern that does not call for emergency services, you can ask to speak to a doctor at any time through Noria Health Hub. This on-demand medical consultation is billed separately from the assessment pathway.
Further reading
The two companion articles.
How” rel=”noopener” target=”_blank”>https://noriahealth.com/blog/how-we-build-a-health-assessment/”>How we build a health assessment, and why certain examinations are left out of it
Waiting” rel=”noopener” target=”_blank”>https://noriahealth.com/blog/what-delay-really-costs-what-the-studies-say/”>Waiting for a diagnosis or a treatment: what the studies actually measure
Sources
Corley DA et al., Adenoma Detection Rate and Risk of Colorectal Cancer and Death, New England Journal of Medicine, 2014. https://doi.org/10.1056/NEJMoa1309086
Zhao S et al., Magnitude, Risk Factors, and Factors Associated With Adenoma Miss Rate of Tandem Colonoscopy, Gastroenterology, 2019. https://pubmed.ncbi.nlm.nih.gov/30738046/
Elmore JG et al., Variability in Interpretive Performance at Screening Mammography, Radiology, 2009. https://doi.org/10.1148/radiol.2533082308
Buist DSM et al., Influence of Annual Interpretive Volume on Screening Mammography Performance, Radiology, 2011. https://doi.org/10.1148/radiol.10101698
COLO-DETECT, Computer-aided detection in colonoscopy, Lancet Gastroenterology and Hepatology, 2024. https://doi.org/10.1016/S2468-1253(24)00161-4
Wallace MB et al., Impact of Artificial Intelligence on Miss Rate of Colorectal Neoplasia, Gastroenterology, 2022. https://doi.org/10.1053/j.gastro.2022.03.007
TIMELY trial, Lancet Gastroenterology and Hepatology, 2024. https://doi.org/10.1016/S2468-1253(24)00187-0
Ladabaum U et al., Computer-aided detection in real-world practice, Gastroenterology, 2023. https://doi.org/10.1053/j.gastro.2022.12.004