When Covid came to Massachusetts, it forced Constance Lehman to change how Massachusetts General Hospital screens women for breast cancer. Many people were skipping regular checkups and scans due to worries about the virus. So the center Lehman codirects began using an artificial intelligence algorithm to predict who is at most risk of developing cancer.
Since the outbreak began, Lehman says, around 20,000 women have skipped routine screening. Normally five of every 1,000 women screened shows signs of cancer. “That’s 100 cancers that we haven’t diagnosed,” she says.
Lehman says the AI approach has helped identify a number of women who, when persuaded to come in for routine screening, turn out to have early signs of cancer. The women flagged by the algorithm were three times as likely to develop cancer; previous statistical techniques were no better than random.
The algorithm analyzes prior mammograms, and seems to work even when physicians did not see warning signs in those earlier scans. “What the AI tools are doing is they’re extracting information that my eye and my brain can’t,” she says.
Researchers have long touted the potential for AI analysis in medical imaging, and some tools have found their way into medical care. Lehman has been working with researchers at MIT for several years on ways to apply AI to cancer screening.
But AI is potentially even more useful as a way to more accurately predict risk. Breast cancer screening sometimes involves not just examining a mammogram for precursors of cancer, but collecting patient information and feeding both into a statistical model to determine the need for follow-up screening.
Adam Yala, a PhD student at MIT, began developing the algorithm Lehman is using, called Mirai, before Covid. He says the goal of using AI is to improve early detection and to reduce the stress and cost of false positives.
To create Mirai, Yala had to overcome problems that have bedeviled other efforts to use AI in radiology. He used an adversarial machine learning approach, where one algorithm tries to deceive another, to account for differences among radiology machines, which could mean that patients that face the same risk of breast cancer get different scores. The model was also designed to aggregate data from several years, making it more accurate than previous efforts that include less data.
The MIT algorithm analyzes the standard four views in a mammogram, from which it then infers information about a patient that is often not collected, such as history of surgery or hormone factors such as menopause. This can help if that data has not been collected by a doctor already. Details of the work are outlined in a paper published today in the journal Science Translational Medicine.
