Artificial intelligence software to detect cancerous polyps will soon be tested during actual colonoscopies.
A colonoscopy may be the best test for finding colorectal cancer, but its true claim to fame is preventing colorectal cancers. The key is the detection rate of precancerous polyps called adenomas, which varies from 7 percent to 53 percent depending on the skill of the colonoscopy practitioner.
For each 1 percent increase in the adenoma detection rate (ADR), the risk of developing colorectal cancer within five years of a colonoscopy is reduced by 3 percent to 6 percent. The most experienced colonoscopists have been able to reduce their patients’ cancer risk by 82 percent.
AI can improve detection rate
Promising new research at UC Irvine indicates that colonoscopies aided by artificial intelligence (AI) software can dramatically improve ADR, even among highly proficient colonoscopists.
The research is being conducted by a team led by UCI Health gastroenterologist Dr. William E. Karnes, who is working with AI specialists at DocBot, a company that began through UCI Applied Innovation.
Essentially, all colorectal cancers start as a benign precancerous polyp, Karnes said. The most common of these are adenomas, which have a mean dwell time of 10-plus years to progress from normal tissue to colorectal cancer. Colonoscopy remains the gold standard for finding adenomas
and is the only test capable of removing all adenomas found in the colon.
Karnes and Professor Pierre Baldi in the UCI Department of Computer Science first developed a proof-of-concept algorithm to spot signs of an adenoma, based on a database of polyp images that Karnes and his gastroenterology team assembled.
The plan is to test it this summer in real time during colonoscopies. The first round of testing involved the use of videos taken during previous colonoscopies and reviewed afterward.
Early results are promising
“We had three experts at UCI with very high ADRs review the videos,” said Karnes, director of the high-risk colon cancer program and of colonoscopy quality at UCI’s H.H. Chao Digestive Disease Center.
“We had them mark every polyp they saw. On first review, they found about 20 percent more polyps than the original colonoscopist had removed. Then they viewed it with an AI overlay; AI found all the polyps they had discovered, plus about 20 percent more polyps than even they did.”
The software operates on a standard desktop machine, processing 98 images per second, which is nearly four times faster than required for live video.
Up next: testing during colonoscopies
The important next step is to test the software during actual colonoscopies.
“We have to make sure the interface is absolutely perfect for real-time colonoscopy,” Karnes said. “Once we’ve achieved that, we’ll be ready to do our randomized study. We’ll see — with the overlay and without it — if the AI increases the rate of polyp discovery.”
If the software operates as expected, it will warn the colonoscopist during the procedure when it identifies tissue showing signs of being a polyp and should be removed. With such assistance, AI could bring all colonoscopists, including those with low ADR rates, to top ADR levels, and ensure that patients are getting a high-detection procedure no matter who their doctors are.
Eventually, Karnes said, the plan is to take the software to the point where it can identify which polyps are benign and don’t call for laboratory analysis and which do.
“It’s expected that hundreds of millions of dollars of medical dollars could be saved if we could diagnose a polyp by looking at it,” he said.