Cloud AI More Than Doubled the Detection Rate for Large Colon Polyps

A cloud-based artificial-intelligence system gave endoscopists a live second set of eyes during colonoscopy—and in a randomized trial across four European countries, it more than doubled the per-procedure detection rate for polyps at least 10 millimeters across.

Among 841 analyzed screening and surveillance patients, doctors using the system found 0.82 adenomas per colonoscopy, compared with 0.62 during standard colonoscopy. That was a 33 percent relative increase. For large polyps, the rates were 0.12 per procedure with AI and 0.05 without it, a rate ratio of 2.36.

The trial did not show that the technology prevents cancer or saves lives. It showed something narrower and still important: real-time AI, delivered through the cloud, helped skilled clinicians notice more clinically significant lesions during the procedure in which they could be removed.

Research references: [1]

A second set of eyes during a difficult search

A colonoscopy asks a clinician to inspect a long, folded, moving surface in real time. Small or subtle lesions can hide behind folds, blend into surrounding tissue or appear only briefly on the monitor. Adenomas and sessile serrated lesions matter because some can progress toward colorectal cancer.

The system tested in the EAGLE trial, called CADDIE, did not diagnose the patient or decide what to remove. Video from the endoscope traveled through the hospital network to an AI model in the cloud. When the model detected an area that might contain a polyp, it returned a green bounding box to the endoscopy screen. The endoscopist remained responsible for examining the tissue, deciding whether to remove it and interpreting the clinical situation.

That division of labor is the breakthrough. The machine watches every frame for patterns; the clinician owns the decision.

Research references: [1] [2]

What the randomized trial found

Researchers randomized patients to standard colonoscopy or colonoscopy assisted by CADDIE. The final analysis included 841 patients treated by 22 endoscopists at eight centers in Italy, Spain, Poland and Germany. Participants were undergoing screening or surveillance rather than investigation prompted by symptoms or a positive fecal immunochemical test.

The co-primary efficacy result was adenomas per colonoscopy. AI assistance increased the average from 0.62 to 0.82, a rate ratio of 1.33 with a 95 percent confidence interval from 1.06 to 1.67. The share of procedures in which at least one adenoma was detected also rose from 35.9 percent to 43.2 percent—an absolute difference of 7.3 percentage points.

The large-lesion result was stronger. Polyps at least 10 millimeters across were detected at a rate of 0.12 per procedure with AI and 0.05 without it, a rate ratio of 2.36 with a 95 percent confidence interval from 1.33 to 4.17.

Detection of sessile serrated lesions also increased: 0.08 per procedure with AI versus 0.03 with standard colonoscopy, a rate ratio of 3.30. The confidence interval was wide—from 1.41 to 7.57—so the comparison favored AI while the precise size of the effect remains uncertain.

More alerts would be a poor trade if they encouraged removal of large amounts of irrelevant tissue. The study therefore measured positive percent agreement: the share of resected tissue confirmed as clinically relevant. The AI arm was noninferior, at 53.9 percent versus 53.4 percent in the standard arm, with a prespecified 15-percentage-point noninferiority margin.

A UK Health Research Authority summary reported no serious harms judged to have been caused by the device. Procedures in which no polyps were found took a similar amount of time in both groups; total procedure time was slightly longer with AI, consistent with more lesions being removed.

Research references: [1] [2]

Why the cloud matters

Most clinical AI stories focus on the model. EAGLE also tested a delivery method.

Earlier computer-aided detection systems commonly depended on high-performance hardware installed beside each endoscopy unit. In EAGLE, a lower-specification computer captured the video feed, sent images to the cloud and returned the visual overlay.

The network path was fast enough for real-time use at the participating sites. Average cloud-network latency was 59.4 milliseconds, and 99.6 percent of the procedure time remained below the study’s 100-millisecond threshold. One procedure had an average above that threshold.

That does not prove every hospital network can support the system, or that cloud delivery lowers total costs. It demonstrates that the architecture worked across multiple hospitals while procedures were underway. If it generalizes, hospitals may be able to update clinical AI without replacing a dedicated processor in every room—provided they can meet the security, reliability, interoperability and regulatory requirements.

Research references: [1]

The evidence is promising and bounded

EAGLE tested detection, not long-term outcomes. It did not measure whether the extra lesions found led to fewer interval cancers, lower colorectal-cancer mortality or better quality of life.

The analysis also needs to be read carefully. Investigators randomized 973 patients and excluded 132 from the reported modified intention-to-treat/per-protocol analysis. Those exclusions included withdrawals before or during procedures, 21 completed procedures later found ineligible because of a positive fecal test or uncleared equipment, and 72 procedures excluded for quality, operator-minimum or other protocol reasons.

The study population and equipment also narrow the inference. Participants were screening or surveillance patients, the endoscopists met a minimum detection standard, and the trial used specified Olympus equipment. Results may differ with symptomatic or positive-FIT patients, other clinicians, other hardware or different networks.

Odin Medical sponsored the study and developed CADDIE. Five authors reported consultancy fees from Odin Medical; additional relationships with Olympus and Medtronic were disclosed. Several conflicts were reported as postdating the study. Those disclosures do not negate the randomized result, but independent replication will matter.

Research references: [1] [2]

A practical model for clinical AI

The most convincing medical AI may not be a system that tries to replace a clinician. It may be a system that helps a clinician perform a demanding task more consistently while leaving judgment and responsibility in human hands.

EAGLE offers that kind of evidence. Across eight centers, a cloud-based detector helped endoscopists find more adenomas, more large polyps and more sessile serrated lesions without reducing the clinical relevance of removed tissue on the study’s co-primary measure.

The next test is larger than detection: broader health systems, more varied patients and equipment, and outcomes people can feel—fewer missed cancers, less intensive treatment and more healthy years of life.

For now, the trial marks a meaningful step. AI did not replace the doctor at the screen. It helped the doctor see more.