New AI safeguard aims to keep cancer-diagnosis tools from making confident mistakes
Researchers at Vanderbilt Health and centers in Hong Kong unveiled TRUECAM, a system that wraps around existing pathology AI to flag slides it should not judge and screen out uninformative tissue. In tests on non-small cell lung cancer images it was more accurate than rival approaches and extended to breast, brain and kidney tissue.
AI is increasingly used in the labs that diagnose cancer, and a team at Vanderbilt Health, working with centers in Hong Kong, has built a safeguard meant to stop it from making confident mistakes. Reported June 23 in Nature Biomedical Engineering, the system, called TRUECAM, works as a wrapper around existing pathology models. It flags inputs that fall outside what a model was trained to read and screens out regions carrying no useful information, such as normal or poorly stained tissue, before the AI makes its call. The team tested it on non-small cell lung cancer slides from two separate research consortia, on deliberately out-of-scope images, and on real cases from Queen Mary Hospital in Hong Kong, and it also held up across breast, brain and kidney tissue. The researchers report it was more accurate than rival approaches while staying fast and inexpensive to run. Doctors will watch whether it performs the same way in everyday clinics.