AI Improves Colonoscopy Diagnostics in Prospective Trial

Pam Harrison

August 14, 2018

A computer trained by artificial intelligence (AI) to diagnose diminutive nonneoplastic polyps located in the distal colon performs well enough to allow endoscopists to safely diagnose while leaving the polyps unresected, a Japanese study indicates.

"To our knowledge, this was the first large-scale prospective study to assess the reliability of real-time use of [computer-aided diagnosis (CAD)] for optical assessment of diminutive colorectal polyps," Yuichi Mori, MD, PhD, from Showa University Northern Yokohama Hospital, Yokohama, Japan, and colleagues observe. "We found that CAD designed for endocytoscopy offers performance levels that meet the clinical threshold for using the diagnose-and-leave strategy for diminutive, nonneoplastic rectosigmoid polyps, they note.

"[W]hich may help improve the cost-effectiveness of colonoscopy."

Their study was published online today in the Annals of Internal Medicine.

Investigators compared the diagnostic performance of real-time CAD with gold-standard pathology of resected polyps among 791 patients undergoing colonoscopy for screening, surveillance, or symptomatic indications. Endoscopists, both expert and nonexpert, first used CAD with narrow-band imaging (CAD-NBI), followed by stained mode (CAD-stained) analysis. The primary end point was whether CAD-stained analysis achieved a 90% or greater negative predictive value (NPV) for diagnosing diminutive (≤5 mm) rectosigmoid adenomas, which is the recommended threshold for a diagnose-and-leave nonneoplastic polyp strategy.

Overall, 466 diminutive polyps were assessed by CAD, of which 250 were rectosigmoid polyps. CAD was able to differentiate neoplastic from nonneoplastic polyps with a 98.1% accuracy rate. Of the 250 rectosigmoid polyps analyzed by CAD, the CAD-NBI analysis had a NPV of between 95.2% and 96.5%, whereas the in CAD-stained analysis, NPVs ranged between 93.7% and 96.4%. In contrast, for polyps in the proximal-to-rectosigmoid colon (n = 216), the NPVs were 60.0% for both types of CAD.

Interestingly, in an ad hoc analysis, the researchers found that CAD had a higher NPV for diminutive rectosigmoid adenomas, at 96.4%, than did either expert endoscopists (91.8%) or nonexperts (86.6%).

"[T]he median time required to obtain the first CAD output was 19...seconds for CAD-NBI and 73...seconds for CAD-stained analysis (P < 0.001)," the researchers observe.

Given that endoscopists detected, on average, two diminutive polyps per patient in the current study, the use of CAD-NBI would take between 60 and 90 additional seconds per colonoscopy, an acceptable amount of time, they suggest, given the high performance value of CAD-NBI.

Promising Findings

In an accompanying editorial, Øyvind Holme, MD, PhD, from the Institute of Health and Society, University of Oslo, Norway, and Lars Aabakken, MD, PhD, from the University of Oslo and Oslo University Hospital Rikshospitalet, Norway, call the findings "promising" and point out several features of the system that make it potentially useful for routine clinical practice.

For example, many endoscopes are already equipped with the same narrow-band imaging system as was used with CAD in the current study, and it is both easily activated by the operator and appears to be as good as staining.

Also, "optical diagnosis of the polyps was delivered quickly: A reliable optical biopsy was obtained in about 35 seconds," they write.

"[P]erhaps most important, the CAD system worked as well in the hands of novices as it did among experts," they add. This has been a major hurdle so far in the ability to diagnose colorectal polyps optically, as reliable results from optical diagnosis have so far only been arrived at by expert endoscopists working in an academic setting.

The editorialists caution, however, that CAD detection of colonic polyps still cannot distinguish between hyperplastic polyps and adenomatous polyps proximal to the sigmoid colon, which is a major limitation to the use of CAD at the moment.

Furthermore, the endocytoscopies used in the current study are not readily available, and the CAD modality does not handle sessile serrated polyps well. Sessile serrated polyps have been identified as an important precursor to colorectal cancer in recent years. Still, '[t]he study...represents

another example of the benefit of big data, a prerequisite for the deep-learning technology used to train the CAD system," Holme and colleagues write. "To err is human, but CAD may help us reduce the frequency of human errors."

Mori reports receiving personal fees from Olympus and grants from The Japan Society for the Promotion of Science, as well as from the Japan Agency for Medical Research and Development. He also reports having a patent licensed to Showa University and Cybernet Systems. Coinvestigators also have a number of conflicts of interest, which are listed in the published report. The editorialists have disclosed no relevant financial relationships.

Ann Intern Med. Published online August 13, 2018. Article extract, Editorial extract

For more news, join us on Facebook and Twitter


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.
Post as: