Computer-assisted colonoscopy driven by artificial intelligence (AI) may increase detection of small polyps, but not colorectal neoplasias or advanced adenomas, two new reports show.
A multicenter, parallel, randomized controlled trial in Spain of patients with a positive fecal immunochemical test (FIT) found that use of AI was not associated with better detection of advanced neoplasias.
Because FIT-positive patients have the highest prevalence of these neoplasias, they offer the best context for investigating the ability of computer-aided detection (CAD) to support their diagnosis, the researchers note.
The results were “surprising,” given previous studies showing benefit of AI devices in the detection of small polyps and adenomas, lead author Carolina Mangas-Sanjuan, MD, PhD, Hospital General Universitario Dr. Balmis, Alicante, Spain, told Medscape Medical News.
Similarly, a systematic review and meta-analysis of trials on use of CAD for polyp detection during colonoscopy found increased detection of adenomas, but not advanced adenomas, and higher rates of unnecessary removal of non-neoplastic polyps.
The studies were published online August 28 in Annals of Internal Medicine.
‘Room for Improvement’
The Spanish trial, led by Mangas-Sanjuan, aimed to assess the contribution of CAD (GI Genius, Medtronic) to colonoscopic detection of advanced colorectal neoplasias, adenomas, serrated polyps, and nonpolypoid and right-sided lesions.
Advanced colorectal neoplasia was defined as an advanced adenoma and/or advanced serrated polyp. An absolute increase of 5% in advanced colorectal neoplasia detection with CAD was considered clinically relevant.
A total of 3213 people (mean age, 61 years; 53% men) with a positive FIT were randomly assigned to colonoscopy with or without CAD.
No significant between-group difference was found in the advanced colorectal neoplasia detection rate: 34.8% with CAD vs 34.6% without (adjusted risk ratio [aRR], 1.01).
Similarly, there was no significant difference in the mean number of advanced colorectal neoplasias detected per colonoscopy: 0.54 with CAD vs 0.52 without (adjusted rate ratio, 1.04).
Nor did the adenoma detection rate differ: 64.2% with CAD vs 62% without (aRR, 1.06).
A subanalysis found small effects for CAD per colonoscopy, including detection of an increased mean number of nonpolypoid lesions (0.56 vs 0.47; adjusted rate ratio, 1.19), proximal adenomas (0.94 vs 0.81; adjusted rate ratio, 1.17), and lesions ≤ 5 mm (polyps in general and adenomas and serrated lesions in particular). CAD also enhanced detection of nonpolypoid lesions ≥ 10 mm (6.8% vs 5.5%; aRR, 1.25).
“The current findings are a snapshot of what these systems can currently offer and what can be expected from them,” the authors write. “Detecting more advanced lesions still lies in the hands of experienced endoscopists who can recognize the lesions and achieve adequate mucosal exposure.”
The results “show that it is necessary to continue improving this technology by using broader image databases to train CAD devices to recognize these types of lesions, and that there is significant room for improvement for this promising technology,” Mangas-Sanjuan told Medscape Medical News.
Weighing Benefits and Harms
In the systemic review and meta-analysis, an international team of investigators searched the literature through February 2023 for randomized controlled trials comparing standard colonoscopy with colonoscopy performed with real-time CAD tools for polyp detection. Twenty-one trials involving 18,232 patients were included.
Benefit outcomes included the adenoma detection rate (ADR) (ie, the proportion of patients with at least one adenoma); number of adenomas detected per colonoscopy; advanced adenoma (≥ 10 mm with high-grade dysplasia and villous histology) detection per colonoscopy; number of serrated lesions detected per colonoscopy; and adenoma miss rate.
Harm outcomes included the number of polypectomies for non-neoplastic lesions per colonoscopy and withdrawal time (how long an endoscopist takes to withdraw their endoscope at the end of a colonoscopy).
The ADR was higher in the CAD group (44% vs 35.9%; relative risk, 1.24; low-certainty evidence), corresponding to a 55% relative reduction in miss rate (risk ratio, 0.45; moderate-certainty evidence).
However, the CAD-detected adenomas were not advanced, and more non-neoplastic polyps were removed in the CAD group (0.52 vs 0.34 per colonoscopy; low-certainty evidence), with a marginal increase in mean inspection time (mean difference, 0.47 minute; moderate-certainty evidence).
“The clinical relevance of our analysis depends on the assumption that ADR is an important outcome of screening colonoscopy,” the authors note. “It has been related inversely with incidence and mortality of CRC [colorectal cancer], which in turn represent critical outcomes. Thus, ADR may be considered a surrogate for CRC prevention.
“On the other hand, the possible harm of unnecessary resections deserves further consideration. Most nonneoplastic polyps are diminutive without increase in the risk for adverse events. However, they can represent an additional cost affecting the cost-effectiveness of the intervention.”
Co-first author Marco Spadaccini, MD, PhD, Humanitas University, Milan, Italy, told Medscape Medical News that although the devices may reduce adenoma miss rates during colonoscopy, “they may increase overdiagnosis and overtreatment of non-neoplastic polyps.”
“The lack of improvement in advanced adenoma detection highlights the need to confirm in a clinical setting the actual potential of CAD systems in reducing the risk of post-colonoscopy CRC and CRC-related mortality,” he said.
Dennis L. Shung, MD, MHS, PhD, of Yale University, New Haven, Connecticut, made several key points in a related editorial. Machine-learning algorithms such as the ones used in these trials “are optimized to perform narrow tasks well,” he writes. But, the specific task of finding polyps “is just one component of colorectal cancer prevention, which is the main goal.
“Community participation in screening programs, optimized bowel preparation regimens, and appropriate surveillance require multilevel interventions. Health systems must balance the cost of implementing and maintaining [CAD] with the costs of other components of colorectal cancer prevention,” he writes.
AI devices do not exist in a silo, but “within a clinical workflow system,” he adds. Therefore, their implementation requires “workflow redesign,” rather than simply relying on CAD to do the work.
Trust is another issue to consider, Shung writes. Lower trust can lead to underutilization of CAD, whereas higher trust can lead to overreliance, with endoscopists failing to verify the CAD findings.
“The current gap between randomized controlled trial performance and real-world performance is concerning but likely reflects both differences in clinician behavior outside of trials and the complexity of real-world clinical environments,” Shung writes. “How algorithmic systems partner with clinicians and how these should be designed and refined across heterogeneous systems and contexts are necessary questions that must be explored to minimize disruption and lead to real-world effectiveness.”
The trial by Mangas-Sanjuan and colleagues was funded by Medtronic. Mangas-Sanjuan declared receiving support for the manuscript and grants from Medtronic, which also lent the CAD devices used in the study. The review by Spadaccini and colleagues was funded by the European Commission Horizon 2020 Marie Skłodowska-Curie Individual Fellowship. Spadaccini declared an ESGE Medtronic AI Research Award. Shung declared no relevant financial relationships.
Follow Marilynn Larkin on X: @MarilynnL