From the AGA Journals

AI-based CADe outperforms high-definition white light in colonoscopy

View on the News

Augment, but not replace

Several randomized trials testing artificial intelligence (AI)–assisted colonoscopy showed improvement in adenoma detection. This study adds to the growing body of evidence that computer-aided detection (CADe) systems for adenoma augment adenoma detection rates, even among highly skilled endoscopists whose baseline ADRs are much higher than the currently recommended threshold for quality colonoscopy (25%).

This study also highlights the usefulness of CADe in aiding detection of sessile serrated lesions (SSL). Recognition of SSL appears to be challenging for trainees and the most likely type of missed large adenomas overall.

Monika Fischer, MD, is an associate professor of medicine at Indiana University, Indianapolis

Dr. Monika Fischer

Given its superior performance, compared with high-definition white light colonoscopy, AI-assisted colonoscopy will likely soon become standard of care. Beyond adenoma detection programs such as CADe, there will be systems to aid with the diagnosis and predict histology such as CADx and other AI programs that evaluate the quality of colon examination by the endoscopist. CADe systems are currently quite expensive but expected to be more affordable as new products become available on the market.

AI-based systems will enhance but will not replace the highly skilled operator. As this study pointed out, despite the superior ADR, adenomas were still missed by CADe. The main reason for this was that the missed polyps were not brought into the visual field by the operator. A combination of a CADe program and a distal attachment mucosa exposure device in the hands of an experienced endoscopists might bring the best results.

Monika Fischer, MD, is an associate professor of medicine at Indiana University, Indianapolis. She reported no relevant conflicts of interest.



An artificial intelligence (AI)–based computer-aided polyp detection (CADe) system missed fewer adenomas, polyps, and sessile serrated lesions and identified more adenomas per colonoscopy than a high-definition white light (HDWL) colonoscopy, according to findings from a randomized study.

While adenoma detection by colonoscopy is associated with a reduced risk of interval colon cancer, detection rates of adenomas vary among physicians. AI approaches, such as machine learning and deep learning, may improve adenoma detection rates during colonoscopy and thus potentially improve outcomes for patients, suggested study authors led by Jeremy R. Glissen Brown, MD, of the Beth Israel Deaconess Medical Center and Harvard Medical School, Boston, who reported their trial findings in Clinical Gastroenterology and Hepatology.

The investigators explained that, although AI approaches may offer benefits in adenoma detection, there have been no prospective data for U.S. populations on the efficacy of an AI-based CADe system for improving adenoma detection rates (ADRs) and reducing adenoma miss rates (AMRs). To overcome this research gap, the investigators performed a prospective, multicenter, single-blind randomized tandem colonoscopy study which assessed a deep learning–based CADe system in 232 patients.

Individuals who presented to the four included U.S. medical centers for either colorectal cancer screening or surveillance were randomly assigned to the CADe system colonoscopy first (n = 116) or HDWL colonoscopy first (n = 116). This was immediately followed by the other procedure, in tandem fashion, performed by the same endoscopist. AMR was the primary outcome of interest, while secondary outcomes were adenomas per colonoscopy (APC) and the miss rate of sessile serrated lesions (SSL).

The researchers excluded 9 patients, which resulted in a total patient population of 223 patients. Approximately 45.3% of the cohort was female, 67.7% were White, and 21% were Black. Most patients (60%) were indicated for primary colorectal cancer screening.

Compared with the HDWL-first group, the AMR was significantly lower in the CADe-first group (31.25% vs. 20.12%, respectively; P = .0247). The researchers commented that, although the CADe system resulted in a statistically significantly lower AMR, the rate still reflects missed adenomas.

Additionally, the CADe-first group had a lower SSL miss rate, compared with the HDWL-first group (7.14% vs. 42.11%, respectively; P = .0482). The researchers noted that their study is one of the first research studies to show that a computer-assisted polyp detection system can reduce the SSL miss rate. The first-pass APC was also significantly higher in the CADe-first group (1.19 vs. 0.90; P = .0323). No statistically significant difference was observed between the groups in regard to the first-pass ADR (50.44% for the CADe-first group vs. 43.64 % for the HDWL-first group; P = .3091).

A multivariate logistic regression analysis identified three significant factors predictive of missed polyps: use of HDWL first vs. the computer-assisted detection system first (odds ratio, 1.8830; P = .0214), age 65 years or younger (OR, 1.7390; P = .0451), and right colon vs. other location (OR, 1.7865; P = .0436).

According to the researchers, the study was not powered to identify differences in ADR, thereby limiting the interpretation of this analysis. In addition, the investigators noted that the tandem colonoscopy study design is limited in its generalizability to real-world clinical settings. Also, given that endoscopists were not blinded to group assignments while performing each withdrawal, the researchers commented that “it is possible that endoscopist performance was influenced by being observed or that endoscopists who participated for the length of the study became over-reliant on” the CADe system during withdrawal, resulting in an underestimate or overestimation of the system’s performance.

The authors concluded that their findings suggest that an AI-based CADe system with colonoscopy “has the potential to decrease interprovider variability in colonoscopy quality by reducing AMR, even in experienced providers.”

This was an investigator-initiated study, with research software and study funding provided by Wision AI. The investigators reported relationships with Wision AI, as well as Olympus, Fujifilm, and Medtronic.

Next Article: