Colonoscopy: a real-time AI helps doctors find more adenomas

Why do lesions go unnoticed during colonoscopy?

Colonoscopy means examining the inside of the colon with a small camera. It lets doctors spot and remove polyps, small growths, some of which, called adenomas, can turn into colorectal cancer over time.

But the authors note that one-quarter of colorectal neoplasias (precancerous or cancerous lesions) are missed during screening colonoscopies. A missed lesion can keep growing and become a cancer.

Better detection tools are therefore of real interest: a tool that helps the doctor notice more of these lesions could make screening colonoscopy more useful.

How was the AI tested?

The researchers ran a randomised trial: a random draw decides whether a patient is examined with or without the AI's help, which lets two similar groups be compared. In all, 685 people were included at 3 Italian centres in 2019. They came for screening, for surveillance after polyp removal, or for a work-up after a positive test for blood in the stool or symptoms.

The AI system, called GI-Genius, analyses the images in real time. When it sees a suspicious area, it shows a green box on the endoscopist's screen. The doctor then decides what to do. Removed lesions were examined in the laboratory, which serves as the reference. A minimum withdrawal time of 6 minutes was required in both groups.

Both groups had the same high-definition equipment and the same minimum examination time, so that the only intended difference was the AI help.

What were the results?

The main outcome was the adenoma detection rate, meaning the share of patients with at least one adenoma or cancer confirmed by laboratory analysis. It was 54.8% with the AI, against 40.4% without (relative risk of 1.30). The average number of adenomas found per colonoscopy was 1.07 with the AI, against 0.71.

Small adenomas of 5 mm or less were found in 33.7% of patients with the AI against 26.5%, and those of 6 to 9 mm in 10.6% against 5.8%. Withdrawal time (417 seconds against 435) and the share of patients having non-tumour lesions removed (26.0% against 28.7%) were not significantly different.

In plain terms, the AI-assisted group had a clearly higher share of patients in whom at least one adenoma was found.

What does this change for patients?

Finding more adenomas during the exam gives a better chance of removing them before they become dangerous. The AI acts as an attentive second pair of eyes, without lengthening the exam and without leading to more unnecessary removals.

This result comes from a randomised trial, a stronger type of study than simple comparisons made after the fact.

The authors conclude that adding the AI significantly increases detection without increasing withdrawal time.

What are the limits of this study?

The trial measures the detection of adenomas, not directly the number of cancers avoided or patient survival. It took place at 3 centres in Italy over a short period (September to November 2019), with a single AI system. Results could differ in other countries or with other teams.

Also, finding more small adenomas does not always mean each one would have turned into cancer. Longer studies are needed to measure the effect on cancers actually prevented.