PHILADELPHIA—An analysis of data from the Minneapolis VA Medical Center showed that colonoscopies assisted by artificial intelligence technology are significantly more likely to result in the resection of benign lesions only. Compared with unassisted colonoscopies, the absolute increase was 4%, according to Tessa Herman, MD, a chief resident at the University of Minnesota, in Minneapolis.

Examining the Impact Of AI on Colonoscopy

AI-assisted colonoscopy (AIAC) with computer-aided detection (CADe) technology is designed to improve colonoscopy quality, primarily by improving the adenoma detection rate (ADR) and reducing the adenoma miss rate. But AIAC has been shown to increase the resection of non-neoplastic lesions, according to data from 12 randomized controlled trials that included approximately 13,000 patients (Ann Intern Med 2023;176[9]:1209-1220).

To examine how often colonoscopies performed for colorectal cancer prevention result in the resection of benign lesions without synchronous adenoma resection, Dr. Herman and her co-investigators conducted an ad hoc analysis of data collected during a prior prospective, single-center pre-/post-implementation trial.

The population included all adults undergoing colonoscopies for screening, surveillance or fecal immunochemical test–positive indications at the Minneapolis VA Medical Center, excluding exams done for diagnostic indications or inflammatory bowel disease, as well as incomplete colonoscopies and those with inadequate bowel preparation. The primary end point was the proportion of colonoscopies resulting in the resection of benign lesions only, without synchronous adenoma resection, for AIAC versus unassisted exams. They defined “benign lesions” as any lesions other than tubular adenomas, sessile serrated lesions or advanced adenomas.

image

In total, 1,040 colonoscopies were performed: 599 with AI assistance and 441 without. The cohorts were balanced, without statistically significant demographic differences. The same eight endoscopists, all naive to AI-assistance technology, were involved throughout the study period.

More Benign Lesions Removed With AI Assistance

The rate of resection of benign non-adenomatous lesions with AI assistance was 30.9%, versus 22.7% without AI assistance, and a significantly higher proportion of benign lesion resections with no synchronous adenomas were removed during AIACs (12.4% vs. 8.4%; P=0.04), Dr. Herman reported at ACG 2024 (abstract 12).

“We found that the most common benign lesions were benign colonic mucosa, lymphoid aggregates and hyperplastic polyps,” she said, adding that data on polyp characteristics such as size were not available. Twice as many benign colonic mucosa (12) and lymphoid aggregates (20) were removed in the AI-assisted group as in the unassisted group (six and 10, respectively). A total of 28 hyperplastic lesions, with or without another benign lesion, were removed in the AI-assisted group compared with 17 in the unassisted group.

“I want to emphasize that not all of these polypectomies were performed inappropriately, as some of these lesions did need to be resected, so that is not necessarily a fault of the AI,” she commented.

“Given that the removal of adenomatous polyps is what reduces colorectal cancer, the value of AIAC is highest when it results in higher ADRs. The value of AIAC is then lowest when AIAC results in only the detection of non-adenomatous lesions. Our analysis found that AIAC resulted in a statistically significant increase in the latter—that is to say, there were more cases where only non-adenomatous lesions were removed,” she said. “Polypectomies in this situation have the potential to increase procedural risk and increase healthcare costs for a multitude of reasons, including supply costs and upcoding from screening to diagnostic colonoscopy.”

Further Discussion

The original prospective trial compared GI Genius (Medtronic) and EndoScreener (Micro-Tech/Wision A.I.). These devices returned similar results in terms of adenoma detection and were moderately comparable in detecting benign lesions.

Dr. Herman acknowledged that hybrid CADe and computer-aided diagnosis (CADx) systems are being refined to better characterize lesions and reduce unnecessary polypectomies, “but there’s a ways to go” in this area, she said, in terms of technology development and clinical adoption.

Vani J.A. Konda, MD, the medical director of the Baylor Scott & White Center for Esophageal Diseases, in Dallas, commented that although AI-assisted colonoscopy has the “exciting potential” to detect more polyps, the study’s findings highlight that “mere quantity is not necessarily the goal” of this technology. Its impact on colon cancer mortality relies on its detection and the removal of specific types of lesions that can become malignant over time. “This study demonstrates the gap that we currently have with AI-assisted colonoscopy,” she said.

This gap can be mitigated with further development of the technology, such as the addition of CADx, which can differentiate among types of polyps needing resection versus those that may be left behind. It would also be helpful to understand which endoscopists may optimize their performance with CADe versus those deriving less benefit, Dr, Konda suggested.

“It is important to remind ourselves that AI is a powerful tool, but ultimately,” she said, “the responsibility still relies on the minds and the hands of the endoscopist to make each clinical decision.”

—Caroline Helwick

Dr. Herman reported no relevant financial disclosures. Dr. Konda has served as a consultant to or an advisor for Ambu, Braintree/Sebela, Castle, Exact Sciences, Medtronic, Olympus and Pentax.

This article is from the December 2024 print issue.