Photo: Faith Ninivaggi
A study published in Clinical Gastroenterology and Hepatology highlights findings from the first-ever in-human trial using a special real-time artificial intelligence system to diagnose cancer in the digestive system.
Led by Neil Marya, MD, assistant professor of medicine, co-director of the Program in Digital Medicine, and director of the Digital Medicine Fellowship at UMass Chan Medical School, the SMART-AI trial showed that using AI was better at diagnosing cholangiocarcinoma, or bile duct cancer, than a biopsy.
Dr. Marya and his team used a tiny camera called a cholangioscope to view the biliary strictures, which are narrowings in the bile ducts that carry bile from the liver and gallbladder to the small intestine. They uploaded an AI that Marya developed into a computer that analyzed the video stream during real-time procedures for suspected cholangiocarcinoma.
“Artificial intelligence could change the way we approach how we diagnose cancer. This study is an important first step,” Marya said. “Artificial intelligence in gastrointestinal disease has mostly been focused on colon polyp detection. This is truly the first AI to be used for the purpose of determining if biliary strictures are benign or malignant.”
“During live patient cases, the AI was run and it showed that it’s better than a biopsy,” Marya added. “It’s telling somebody if they have cancer or not.”
A total of 41 patients were enrolled in the trial and had biliary strictures analyzed by the cholangioscopy AI. The trial results showed that the AI had greater accuracy in classifying the strictures than both standard sampling techniques (87.8 percent accuracy versus 67.4 percent) and by visual appearance from experienced endoscopists (87.8 percent versus 63.1 percent).
To develop the AI, Marya and his team first used cholangioscopy videos from previous procedures at UMass Memorial Health and the Mayo Clinic and trained the AI system to mimic a human assessment of the video to tell them whether the patient had cholangiocarcinoma.
“What we’ve seen in other studies is that when physicians look at images or videos, they’re actually very good at telling us if it’s cancer or not just based on the visual appearance,” Marya said. “Our idea was if we could train an AI that could mimic our ability to make a visual impression, maybe we could develop a quantitative tool that could tell us the likelihood that somebody may have cancer just based on a visual appearance. So that’s what we did.”
The study was funded through a Prize for Academic Collaboration and Excellence award from UMass Memorial Medical Group in partnership with UMass Chan; the Mayo Clinic MAX Innovation award in Gastroenterology and Hepatology; and an Acorn Innovation Award from MassVentures.
