- Design
- Systematic review and network meta-analysis of randomised trials, GRADE assessed
- Population
- 13 trials, 4,156 adults undergoing colonoscopy, 8 AI systems
- Primary outcome
- Polyps detected per colonoscopy, by size
- Effect
- ≤5 mm SMD 0.21 (0.07 to 0.35); 6–9 mm 0.02; ≥10 mm 0.01
A network meta-analysis pooled 13 randomised trials (4,156 participants) comparing eight AI detection systems with standard colonoscopy, stratified by polyp size. There were no head-to-head AI comparisons.
For polyps of 5 mm or less, AI showed a modest advantage (standardised mean difference 0.21), with high heterogeneity and a prediction interval crossing no effect. For 6–9 mm and 10 mm or larger, effects were minimal (0.02 and 0.01). Certainty was very low for diminutive polyps and low for larger ones. No platform was clearly superior.
The benefit of computer-aided detection appears to come mostly from small polyps, which carry the lowest cancer risk. It is a useful adjunct, but it is not a substitute for careful withdrawal technique and good bowel preparation.
- Keep withdrawal time and bowel preparation quality as the core of detection
- Expect AI to raise counts of small polyps rather than find more large lesions
- Plan for more diminutive polypectomies and pathology when AI is introduced
- Audit adenoma detection rate before and after adoption
Why it matters
It tempers the expectation that AI will find the larger lesions that matter most for cancer prevention.
Don't overread it
Certainty was very low to low, and comparisons between AI systems were entirely indirect.
The statistics, in plain English
A standardised mean difference of 0.21 is a small effect. The prediction interval, from −1.12 to 1.54, shows what a new setting might see, and it includes no benefit. I² of 87% means trials disagreed substantially. Rankings between systems from indirect comparisons are exploratory only.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for gastroenterology & hepatology, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free