Higher rates of colorectal polyp and adenoma detection were seen with a real-time computer-assisted detection (CADe) system based on deep learning compared with conventional physician interpretation, a randomized study in China found.
In the unblinded study, the artificial intelligence (AI) system significantly increased the adenoma detection rate (ADR) by about 50% (29.1% vs 20.3%, P<0.001), Xiaogang Liu, MD, and colleagues at Sichuan Provincial People's Hospital in Chengdu wrote in Gut.
It also increased the mean number of adenomas detected per patient: 0.53 vs 0.31 (P<0.001). This increase was due, not surprisingly, to a higher number of diminutive adenomas found (185 vs 102, P<0.001). As well, there was a non-significant trend toward greater detection of larger adenomas (77 vs 58, P=0.075).
AI also significantly increased the number of detected hyperplastic polyps (114 vs 52, P<0.001), the researchers found.
Automatic detection of these lesions could help to achieve high stable ADRs in future practice, the researchers stated. “Given its high accuracy, fidelity and stability, the current CADe system is potentially applicable in current clinical practice for better detection of colon polyps,” Liu and colleagues wrote. They cautioned, however, that clinical outcomes remain unclear, and further improvements such as polyp differentiation are needed.
Colorectal ADR is considered a quality indicator of screening colonoscopy and has been shown to correlate with reductions in interval cancers. Detection rate is understood to be the proportion of colonoscopy patients having at least one lesion identified.
The current trial, the first randomized study of its kind, assigned consecutive patients undergoing colonoscopy at the hospital during 2017-2018 to undergo diagnostic colonoscopy with or without CADe. The AI system delivered a sound alert and simultaneous visual notice when it flagged a lesion.
Of the 1,058 patients included, 536 received standard colonoscopy and 522 had colonoscopy with CADe. Mean patient age was about 50, and mean BMI was about 23. Half were women. The mean total withdrawal time was a few seconds longer in the experimental arm (6.89 minutes vs 6.39 minutes), and CADe resulted in 229 more biopsies than unaided colonoscopy.
The mean number of polyps detected per colonoscopy in the control versus the intervention group was 0.51 and 0.97 (P<0.001). Polyp detection rates for the control and CAD groups were 0.29 and 0.45, respectively (odds ratio 2.0, 95% CI 1.5-2.5, P<0.001).
Findings were similar for adenomas (P<0.001 for all comparisons):
- Mean number detected per colonoscopy: 0.31 control vs 0.53 CADe
- Detection rate: 0.20 control vs 0.29 CADe (OR 1.61, 95% CI 1.21-2.14)
Most CAD-detected adenomas, however, were diminutive and likely to be missed within the visual field. Still, the authors argued, although small adenomas confer less risk for malignancy, the increase in overall adenoma detection rate may eventually contribute to a decreased risk of interval colorectal cancers.
In total, there were 39 false positives in the intervention group, averaging 0.075 false alarms per procedure. These were due to local bleeding, medication capsule, and undigested debris. (Study authors did not specify how false positives were identified, whether through repeat colonoscopy or some other method.)
Describing the study as “well done,” and commending the randomized design, Thomas F. Imperiale, MD, of Indiana University School of Medicine in Indianapolis, told MedPage Today the findings are not surprising.
“AI does help improve adenoma detection. However, here it only improved diminutive adenomas of 5 mm or smaller, most of which will not progress to colorectal cancer. Therefore the clinical benefits are unclear,” said Imperiale, who was not involved with the study. “More trials in higher-risk populations would be useful.”
Last year, MedPage Today reported on Japanese research in which CAD showed promise for real-time differentiation between neoplastic polyps that require resection and non-neoplastic polyps that can be left in place.
Among study limitations, the authors cited the unexpected false positives, the unblinded design, and the lack of external validity and applicability to other populations with higher baseline ADRs.
In addition, the study did not control for the fatigue level of participating endoscopists. And because of the inadequate sample size of colonoscopies performed by junior endoscopists, further studies are needed to show the effectiveness of this CAD system at different levels of training. Lastly, the study was conducted using Olympus colonoscopy equipment, and the AI system may not be adaptable to other manufacturers’ equipment.
This study received no specific funding. The CAD system was provided free of charge by Shanghai Wision AI Co., Ltd. The authors disclosed no competing interests. Imperiale and Allison reported no competing interests in relation to their comments.