Dr. William Karnes’ directive when he arrived at UCI Health was nothing less than wiping out colorectal cancer in Orange County.
Although it is the second leading cause of U.S. cancer deaths, colorectal cancer is highly curable — even preventable — with early detection during regular screenings.
The veteran gastroenterologist assessed his three-prong challenge:
"If everyone of screening age got a colonoscopy and every physician was doing the procedure well, we could reduce the risk by 90% nationally," Karnes says. "But people think it's an embarrassing procedure and the colon prep isn't fun. It's an easy test for people to put off."
Gathering the data
First Karnes set to work improving UCI Health colonoscopy detection rates. He began by amassing an impressive database of images collected during colonoscopies.
Within a year, the tens of thousands of images helped him and fellow gastroenterologists at the H.H. Chao Comprehensive Digestive Disease Center (CDDC) increase their already good overall detection rates by 50%.
To improve further, Karnes tapped the power of artificial intelligence (AI) with Andrew Ninh, an information technology expert who was part of the Wayfinder incubator program at UCI Beall Applied Innovation.
They matched hundreds of thousands of images — the world’s largest data set of benign, precancerous and confirmed malignant polyps — with deep-learning algorithms to develop a smarter, faster scan.
AI Polyp Detector: ‘A game-changer’
The result is the AI Polyp Detector, developed by Ninh and Karnes at Docbot, a medical informatics startup they co-founded at UCI. When used during a colonoscopy, the software gives physicians real-time analysis that dramatically increases their ability to spot and remove polyps — with more than 96% accuracy.
"It's a game-changer," Karnes says of the AI software, which is incorporated with the camera device he uses to inspect the colon and small intestine during a colonoscopy.
"It takes about 10 milliseconds to process an image, which is four times faster than it needs to be run in real time. The doctor gets this feedback while watching the screen during a colonoscopy. It can spot a flat polyp that otherwise may be missed."
Detection is key
All colorectal cancers begin as a benign polyp. The most common are adenomas, which take about 10 years to progress into colorectal cancer. With regular colonoscopies, these precancerous adenomas, can be identified and removed.
But the adenoma detection rate — or ADR — varies widely among colonoscopists. The AI polyp detector could ensure that all patients get a high-detection procedure no matter who their doctors are, Karnes says.
There is also potential for significant cost savings.
Instant analysis of polyps
Karnes and fellow UCI Health gastroenterologists using the AI software now send polyps they have removed to a pathologist for testing. It takes about two weeks for the lab to officially verify a polyp’s pathology.
The AI tool now can predict on the spot whether a polyp is precancerous or malignant with 94% accuracy.
With further refinements in software accuracy and widespread adoption of the technology, such testing for anything other than malignant polyps could save an estimated $1 billion in lab costs nationally, Karnes and Ninh say.
Making colon cancer obsolete
Meanwhile, he and other CDDC gastroenterologists are continuing to improve the AI algorithms with each colonoscopy they perform.
Karnes also is tackling another one of those obstacles to his goal: resistance to get getting a colonoscopy in the first place.
He has adopted a way to make them more comfortable, using water to help float the scope through the twists and turns of the intestines. The water bath has the added benefit of making hard-to-find flat polyps easier to see and remove.
And his colleague and UCI Health gastroenterologist Dr. Jason Samarasena, who is also part of Docbot’s leadership team, is testing a low-residue diet that may well make fasting and those colon prep drinks obsolete.
Wiping out colorectal cancer in Orange County may not be such a tall order after all.