High-performing colonoscopists can identify 20% more polyps with artificial intelligence (AI) than without, a new study indicates.
Some gastroenterologists have expressed concern, worrying that they will be put out of work, “but I don’t see that happening,” said William Karnes, MD, from UC Irvine Health in California.
“I see what we’re doing with AI as primarily a way to make the colonoscopists’s job easier, giving more time to our patients and less time to paperwork,” he explained.
When Karnes and his colleagues applied a video-based machine-learning model to about 4100 images with polyps and 4500 without, the accuracy rate for polyp detection was 96%, he reported here at Digestive Disease Week 2018.
Four colonoscopists with career adenoma detection rates of at least 50% then examined the videos with and without the AI video overlay.
The researchers hope that the use of AI will ultimately lead to fewer cases of interval colorectal cancer, because 50% to 80% are the result of missed polyps.
“This was basically a feasibility study to try to mimic what happens during a colonoscopy, using some deidentified videos from older studies, to see whether AI enhanced our ability to see polyps,” Karnes told Medscape Medical News.
“When I saw the first overlay on video, I thought, ‘this is so cool’,” he told Medscape Medical News. “I was actually blown away.”
The 8641 images used in the study were considered diverse and balanced, and included all portions of the colon plus random features, such as debris and diverticuli. The machine-learning model, which incorporated a so-called convolutional neural network, operated at a processing speed exceeding that required for live video.
Using the AI polyp detection technique, the four colonoscopists were tasked with identifying all polyps in nine colonoscopy videos. Karnes, who has performed more than 20,000 colonoscopies, was the only one of the four to review machine-learning-overlaid videos, and he determined the uniqueness of each polyp and the confidence level — low to high — of true polyp presence.
This limitation could create potential bias, he acknowledged, but it will be remedied in upcoming clinical trials.
I was actually blown away.
Karnes compared the AI overlay video with a video stream brought into a small gaming computer. Before sending the video to the monitor, an algorithm assessed each frame to determine if a polyp was present and, if so, placed a box around the polyp.
“So we’re seeing, in real time, the original video along with a little green box if a polyp is present,” he explained.
High Performance
No unique polyps were missed with the machine-learning model. In fact, with the machine-overlaid videos, the number of unique polyps detected increased from 36 to 45. Confidence was high for three of the additional nine polyps detected, and low for six.
The AI model produced a false-positive rate of only 7%, and a negative predictive value of 99.5%, Karnes reported. This means that gastroenterologists using this technology might not have to wait for pathology results.
“When combined with accurate optical pathology, AI may eventually bring us to real-time measures of adenoma detection rate,” he said.
The technology, when used by gastroenterologists during colonoscopy, will likely prompt more careful inspection and the detection of additional polyps, Karnes noted. Multiple centers are already lined up to conduct primarily randomized clinical trials on the technique, he said, and data are expected to accrue quickly thereafter.
AI is already being used in other facets of colonoscopy. It can assess the thoroughness of bowel prep, for example, and monitor when the scope is placed inside and removed from the colon.
After conversations with practicing gastroenterologists, Karnes said he believes most will be “very eager” to try this technique.
AI approaches to polyp detection are “what’s going to happen,” said Daniel Stonewall Anderson, MD, from Kaiser Permanente in San Diego, who is president of the California Colorectal Cancer Coalition. However, he noted, the results “seem very, very early.”
“The AI looking at the images will eventually help us,” he told Medscape Medical News. “It is interesting that removing these tiny polyps that would never mean anything [in terms of cancer development] would come to mean something because they’re an index of quality. It means we’re looking well.”
Karnes is a cofounder of DocBot, a spinoff of this study that might eventually lead to profits for the University of California, Irvine. Anderson has disclosed no relevant financial relationships.
Digestive Disease Week (DDW) 2018: Presentation 133. Presented June 2, 2018.
