Michael Wallace a has performed hundreds of colonoscopies during his 20 years as a gastroenterologist. He thinks he’s good enough at recognizing growths, or polyps, that can arise along the ridges of the colon and potentially turn into cancer. But it’s not always perfect. Sometimes polyps are flat and difficult to see. Other times, the doctors miss them. “We’re all humans,” says Wallace, who works at the Mayo Clinic. After a morning of back-to-back procedures that require attention to the smallest detail, he says, “we’re getting tired.”
Colonoscopies, if unpleasant, are very effective to identify precancerous polyps and prevent colon cancer. But the effectiveness of the procedure largely depends on the abilities of the doctor performing it. Now the Food and Drug Administration has approved a new tool that promises to help doctors recognize precancerous growths during a colonoscopy: a artificial intelligence system manufactured by Medtronic. Doctors say, along with other measures, the tool could help improve diagnoses. “We really have the opportunity to completely rule out colon cancer in anyone who is screened,” says Wallace, who consulted Medtronic on the project.
The Medtronic system, called GI Genius, has seen more colon inside than most doctors. Medtronic and its partner Cosmo Pharmaceuticals trained the polyp recognition algorithm by reviewing more than 13 million videos of colonoscopies performed in Europe and the United States that Cosmo had collected during drug trials. To “train” the AI to distinguish potentially dangerous growths, the images were labeled by gastroenterologists as normal or unhealthy tissue. Then the AI was tested on polyps that were increasingly difficult to recognize, starting with colonoscopies performed under perfect conditions and moving on to more difficult challenges, like distinguishing a polyp that was very small, only in range. camera briefly, or hidden in a dark spot.
The system, which can be added to litters doctors already use to perform colonoscopy, tracks when the doctor probes the colon, highlighting potential polyps with a green box. GI Genius was approved in Europe in October 2019 and is the first AI authorized by the FDA to help detect colorectal polyps. “He found things that even I missed,” says Wallace, co-author of the first validation study by GI Genius. “It’s an awesome system.”
Mark Pochapin, a gastroenterologist at NYU Langone who was not involved in the creation of GI Genius, says it makes sense that AI is effective at recognizing polyps. “There’s less diversity when you look at polyps,” says Pochapin. The millions of colonoscopy videos provide a lot of data to make the algorithm complete. This should protect the system from concerns about bias in other healthcare algorithms. “There are only a limited number of varieties of polyps,” he says.
Medtronic considers GI Genius and other AI tools to be the cornerstone of its future activities, says Giovanni Di Napoli, president of the GI business at Medtronic. To this end, the company has invested a lot of time and resources to gain FDA approval for this device. “It took us almost a year to get FDA approval,” says Di Napoli. “It is not easy.”
Medtronic has sought FDA clearance in what the agency calls its de novo route, which requires applicants to provide information about the safety and efficacy of new devices, including clinical data. This is a longer and more complex application that some other AI medical devices have avoided. Most AI and machine learning medical devices are brought to market using a simplified FDA application known as the 510 (k) path, which only requires them to prove that their devices are similar to other tools already in use and typically takes about six months. According to a study published in The Lancet, of the 222 AI devices released in the US between 2015 and 2020, 92% did so via the 510 (k).