Does AI-Assisted Colonoscopy Help Find Advanced Neoplasia?

A new meta-analysis confirms that use of computer-aided detection (CADe) systems during colonoscopy finds more polyps and adenomas than conventional colonoscopy, but the effect on detection of advanced colorectal neoplasia (ACN) remains unclear.

The research team, which focused on advanced neoplasia due to their clinical importance, found a small increase in the ACN detection rate with CADe but no difference in ACNs detected per colonoscopy.

For ACN, “the ones we really care about, the findings were discordant,” Dennis Shung, MD, MHS, PhD, with Yale School of Medicine, New Haven, Connecticut, said in an interview.

There was a “small positive signal” indicating the potential of AI to detect advanced neoplasia. “However, we can’t say for sure that it will help you find advanced colonic neoplasia,” Shung said.

Jeremy Glissen Brown, MD, MSc, gastroenterologist with Duke Health, Durham, North Carolina, who wasn’t involved in the study, said in an interview that it’s “one of the most comprehensive systematic reviews and meta-analyses to date, examining both parallel and tandem randomized clinical trials (RCTs) of CADe in colonoscopy.”

“The results are generally consistent with prior RCTs and meta-analyses and show an improvement in important quality metrics, mainly an increase in adenomas per colonoscopy (APC), an increase in adenoma detection rate (ADR), and a decrease in adenoma miss rate (AMR),” Glissen Brown noted.

The analysis was published online on October 21 in Annals of Internal Medicine.

Larger, More Accurate Analysis

Prior meta-analyses of AI-assisted colonoscopy included up to 33 RCTs. In their updated meta-analysis, Shung and colleagues included 44 RCTs with 36,201 cases.

“The large sample size included in this study allowed us to examine efficacy of CADe in diagnosis of clinically relevant colonic lesions more accurately, which was not feasible in prior RCTs and reviews given lower sample size,” they wrote.

For polyp detection, CADe-enhanced colonoscopy outperformed conventional colonoscopy in the average number of polyps detected per colonoscopy (1.59 vs 1.27; incidence rate difference [IRD], 0.35) and polyp detection rate (54% vs 46.5%; rate ratio [RR], 1.21).

The same held true for adenoma detection. CADe-enhanced colonoscopy had a higher average APC (0.98 vs 0.78; IRD, 0.22) and ADR (44.7% vs 36.7%; RR, 1.21), coupled with a lower AMR (16.1% vs 35.3%; RR, 0.47).

Average ACN per colonoscopy was similar with and without CADe enhancement (0.16 vs 0.15; IRD, 0.01), but there was a small increase in the ACN detection rate (12.7% vs 11.5%; RR, 1.16).

Results of a subgroup analysis suggest decreased benefit of CADe in patients with positive fecal immunochemical test results, “which may indicate an attenuated benefit for the use of CADe systems for regular screening practice,” the study team wrote.

In a sensitivity analysis of overall adenoma detection according to baseline ADR, there was an increase in the benefit of CADe systems among providers with a lower ADR.

Use of CADe systems led to resection of nearly two extra nonneoplastic polyps per 10 colonoscopies and a “marginal” increase in withdrawal time (0.53 minutes) that may have “limited clinical significance,” the authors noted.

There were no clear differences in performance between the different CADe systems used in the included studies.

All studies were rated as “high concern” for overall bias. Other limitations include study heterogeneity, absence of blinding between conventional and CADe-enhanced colonoscopies, and unaccountable confounding factors.

Ready for Prime Time?

Is routine adoption of AI-assisted colonoscopy ready for prime time? Glissen Brown thinks so, with some caveats.

“We are at a pivotal point in the examination CADe for routine use in colonoscopy. CADe has been ready for prime time in the United States since at least 2021, and the study of CADe has made the field of gastroenterology a clinical leader when it comes to the number of high-quality randomized trials examining AI interventions,” Glissen Brown said.

However, outside of the clinical trial setting, questions about successful deployment and implementation remain, he said.

“These include but are in no way limited to ways to optimize the AI-human interaction in order to produce a successful AI-provider partnership, issues of reimbursement and cost, and issues of ethical AI development and deployment,” Glissen Brown said.

“We also need to continue to assess methods of estimating CADe use on the downstream outcomes that matter, such as the effects of CADe on reducing rates of colon cancer, rates of post-colonoscopy colorectal cancer (CRC), and the effect that CADe might have on CRC-related mortality. In addition, more studies on the patient voice and patient preference as it relates to AI use are greatly needed,” Glissen Brown said.

The American Gastroenterological Association has drafted recommendations on the use of CADe systems during colonoscopy.

Clinicians are invited to review the draft guideline and share feedback during the public comment period, which ends on October 28.

The study had no specific funding. Disclosures for study authors are available with the original article. Glissen Brown is a consultant for Medtronic, Olympus, and Odin Vision. He was also the lead author on one of the studies included in the meta-analysis.