Women with abnormal mammograms often have to wait for weeks to find out whether they have breast cancer.
Now, researchers at UC San Francisco and UC Berkeley have found a way to help reduce the wait and the worry by using artificial intelligence (AI) to quickly identify those who are most likely to have the disease. Once these patients are triaged, they can go through the diagnostic process immediately — from imaging to evaluation and sometimes even biopsy — in a single day.
“This is a really an exciting time,” said Maggie Chung, M.D., the first author of the study, which published May 19 in Nature Digital Medicine. “This moves us closer to personalized care, where we can tailor a plan so that each patient gets the right intervention at the right time.”
Researchers used an open-source AI model called Mirai, which was developed by the study’s senior author, UC Berkeley data scientist Adam Yala, Ph.D. After being trained on hundreds of thousands of mammograms that were linked to patients’ cancer outcomes, the model can recognize subtle patterns in a screening mammogram and predict a woman’s cancer risk in a more powerful way than a physician working alone.
Same-day mammogram analysis
Chung and Yala applied the AI model to more than 4,100 screening mammograms at Zuckerberg San Francisco General Hospital and Trauma Center. It determined that 525 women — about 12.7% of screened patients — were high risk.
Those patients could get an interpretation of their mammograms right after having them done and have additional diagnostic imaging for any suspicious areas the same day. Some of the women who needed biopsies were able to have that done the same day as well.
The model reduced the wait time for a diagnostic evaluation from several weeks to about an hour. And for those who were ultimately diagnosed with breast cancer, it reduced the average wait for a biopsy from more than two months to fewer than 10 days.
The AI model does not replace radiologists or make diagnoses on its own. Instead, it’s a triage tool that helps physicians identify the patients who can benefit most from accelerated care.
“This is a powerful example of how AI can be a collaborative partner for physicians,” said Yala, who, along with Chung, is an assistant professor in the UCSF-UC Berkeley Joint Program in Computational Precision Health. “It shows how we can improve care when we bring clinicians and data scientists together to design these systems.”
The researchers first tested the model on more than 114,000 archival mammograms to ensure it would capture enough high-risk patients without overloading the clinic with too many expedited evaluations. They hope that AI will foster a more tailored approach to breast cancer screening that is personalized to each patients’ breast cancer risk.
“Right now, many women follow the same screening schedule, but their individual risk can be very different,” Chung said. “AI risk assessment gives us the chance to identify the women most likely to benefit from expedited care and get them what they need.”
Authors: Additional authors of the study are Eric Davis, Heather Greenwood, Jessica Hayward, Shinn-Huey Shirley Chou, Bonnie Joe, Loretta Strachowski, Tatiana Kelil, Rita Freimanis, Elissa Price, Kimberly Ray, Amie Lee, all of UCSF.
Funding: The study was funded in part by the National Institutes of Health (grant R37 CA289821 and a T32 training grant).
