Polyp-detection AI moved to the cloud. The lag was 59 milliseconds; the detection tripled.

Computer-aided detection during colonoscopy has a hardware problem and an evidence problem, and a trial published on 26 December takes a run at both.

The hardware problem: existing CADe systems have required on-site high-performance installations — field-programmable gate arrays or GPUs — sitting in the endoscopy suite. That ties hospitals to legacy equipment and makes algorithm updates a procurement exercise [s1].

The evidence problem is larger. CADe reliably finds more small polyps, and it is not clear that finding them helps.

What EAGLE did

The trial, published in npj Digital Medicine, was a parallel-group randomised controlled trial of a real-time cloud-deployed CADe system. Instead of processing video locally, the endoscopy unit streams data to the cloud for analysis, which allows low-specification hardware in the room and frequent algorithm updates [s1].

Crucially, the model was trained on an enhanced dataset weighted toward clinically significant lesions: large polyps of 10 mm or more, and sessile serrated lesions [s1]. That is a deliberate correction of the field's usual failure mode, which is systems tuned to spot the easy small stuff.

841 patients across eight centres in four European countries, with 22 endoscopists, were randomised to standard or CADe-assisted colonoscopy [s1]. It is registered as NCT05730192 [s1].

The co-primary endpoints were superiority on adenomas per colonoscopy and non-inferiority on positive percent agreement — the proportion of resections confirmed to be clinically relevant polyps, which is the endpoint that guards against the system simply prompting more unnecessary removals [s1].

The results

Adenomas per colonoscopy: 0.82 with CADe versus 0.62 without, a ratio of 1.33 (95% CI 1.06–1.67) [s1].

Adenoma detection rate: 43.2% versus 35.9% [s1].

Sessile serrated lesions: 0.08 versus 0.03 per colonoscopy, a ratio of 3.30 (95% CI 1.41–7.57) [s1].

Large polyps of 10 mm or more: 0.12 versus 0.05, a ratio of 2.36 (95% CI 1.33–4.17) [s1].

All improvements were reported at p<0.05, and positive percent agreement was non-inferior [s1].

On the engineering question, average cloud-network latency was 59.4 ms per minute, with 99.6% of measurements under the 100 ms threshold the authors set as the requirement for real-time use [s1].

What the serrated number means, and why the interval is wide

Sessile serrated lesions matter disproportionately. They are flat, pale, easy to miss on withdrawal, and they account for a substantial share of interval colorectal cancers — the cancers that appear between screening rounds in people who had a clean colonoscopy.

A 3.3-fold ratio for detecting them is the most clinically interesting number in this trial. It is also the least precisely estimated. The confidence interval runs from 1.41 to 7.57 [s1], because the absolute counts are small: 0.08 versus 0.03 lesions per colonoscopy across 841 patients. The direction is clear; the magnitude is not.

The same caution applies to large polyps, where the interval spans 1.33 to 4.17 [s1].

The guideline that declined to recommend this

Ten months earlier, the American Gastroenterological Association published a living clinical practice guideline on CADe-assisted colonoscopy and concluded that no recommendation could be made for or against its use, citing very low certainty of evidence for the critical outcomes [s2].

The panel's modelling explains the hesitation. It estimated 11 fewer colorectal cancers and two fewer colorectal cancer deaths per 10,000 individuals, set against 635 more intensive surveillance colonoscopies per 10,000, plus cost and resource implications [s2]. The detection gains were not disputed — an 8% increase in adenoma detection rate (95% CI 6–10%) and a 2% increase in advanced adenoma and/or sessile serrated lesion detection rate (95% CI 0–4%) [s2].

That 2% figure, with a lower bound of zero, is precisely the gap EAGLE is aiming at. A system trained on clinically significant lesions and reporting a tripling in serrated detection is a direct response to the guideline panel's stated reason for withholding a recommendation.

Whether it settles anything is another question. The AGA's concern was never detection; it was the surveillance burden generated downstream. EAGLE's non-inferiority on positive percent agreement addresses part of that — the resections it prompts are, on the trial's measure, appropriate ones [s1] — but a trial of 841 patients cannot report interval cancers or surveillance colonoscopy volumes over a decade, which is the ledger the guideline was balancing.

The rest of the field is still refining the setup

A systematic review and meta-analysis published in Surgical Endoscopy on 18 December illustrates how much of the current CADe literature is about configuration rather than outcomes. It pooled three randomised trials totalling 2,404 patients comparing CADe combined with a mucosal-exposure device against CADe alone [s3].

Adding the device improved adenoma detection rate (RR 1.12, 95% CI 1.03–1.21) and adenomas per colonoscopy (mean difference 0.15, 95% CI 0.06–0.24), without extending caecal intubation or withdrawal time [s3]. It did not improve advanced adenoma detection rate or the number of advanced adenomas per colonoscopy [s3].

A trial sequential analysis found that for adenoma detection rate the cumulative z-line crossed the boundary for effect and came within 43 patients of the required sample size, and for adenomas per colonoscopy crossed the boundary and reached it [s3].

That is a well-executed synthesis of an incremental question, and it lands in the same place as everything else: more adenomas, no signal on the advanced lesions that drive cancer risk.

Why it matters

Two things in EAGLE are genuinely new. Cloud deployment decouples the AI from the hardware in the room, which changes the economics of updating a model and the barrier to entry for smaller units [s1]. And training on clinically significant lesions rather than on whatever the dataset happened to contain is the field admitting that a higher adenoma detection rate is not automatically a better colonoscopy.

Neither is a demonstration that patients live longer. The chain from serrated lesion detected to cancer prevented remains modelled, not measured, and the AGA's arithmetic is the reminder of how much surveillance sits between the two [s2].

What to watch

Whether the serrated-lesion finding replicates in a larger trial with a tighter confidence interval, and whether any registry follows CADe-screened cohorts long enough to report interval cancers. The second is what would move a guideline.

This article is informational and is not medical advice.

Sources

• [s1] Kader R, Hassan C, Lanas Á, et al. "A novel cloud-based artificial intelligence for real-time detection of colorectal neoplasia – a randomized controlled trial (EAGLE)." npj Digital Medicine, 9(1), published online 26 December 2025. https://doi.org/10.1038/s41746-025-02270-1
• [s2] Sultan S, Shung DL, Kolb JM, et al. "AGA Living Clinical Practice Guideline on Computer-Aided Detection-Assisted Colonoscopy." Gastroenterology, 168(4), published online 20 March 2025. https://doi.org/10.1053/j.gastro.2025.01.002
• [s3] Meine GC, Holanda EU, Santo P, et al. "Computer-aided detection with or without mucosal-exposure devices in colonoscopy: a systematic review and meta-analysis with trial sequential analysis." Surgical Endoscopy, published online 18 December 2025. https://doi.org/10.1007/s00464-025-12492-9