Mammography Study Shows Dynamic Changes of AI Risk Scores Years Before Breast Cancer Diagnosis

A large multicenter retrospective cohort study published in Radiology has found that AI-derived risk scores from screening mammograms diverge significantly between women who go on to develop breast cancer and those who remain cancer-free — and this divergence is detectable years before diagnosis.

For the study, researchers applied the deep learning model Mirai to 158,807 screening mammograms drawn from 54,014 women (median age of 61), including 817 women who developed breast cancer.

The study authors found that over a six-year period, the median AI risk score for breast cancer with the Mirai model rose from 2.1 at baseline to 6.6 at the index examination for those with breast cancer. The median AI risk score ranged between 1.8 to 2.2 for women who remained cancer-free during the study period, according to the study.

In a recent interview with Diagnostic Imaging, lead study author Constance Lehman, MD, PhD, emphasized the ability of AI to help “realize the full promise” of screening mammograms in facilitating dynamic risk stratification. While noting the recently established AI risk score threshold of 1.7 percent for five-year risk, Dr. Lehman discussed an example of two women who may both have a 1.6 five-year risk in which dynamic risk assessment can make a key difference.

“Technically, they're under that threshold of 1.7 percent five-year risk (threshold) … adopted by the NCCN and the USPSTF. But if those two women, if one of them was 1.1 last year and the other was 1.6 last year, I know that's different as far as predicting what is (their) risk in the coming five years,” posited Dr. Lehman, the founder/CEO of Clairity and a professor of radiology currently on leave from Harvard Medical School.

In subgroup analyses, the researchers also found that the divergent trajectories of risk scores between those who developed breast cancer and those who did not remained statistically significant regardless of differences in age or breast density.

"We found that the findings overall were very robust in our subgroups, and so this is really something that we can use throughout the cycle of the woman's life, while she's in that phase of screening with screening mammography," noted Dr. Lehman.

Dr. Lehman suggested that a dynamic risk score for screening mammography may lead to other possibilities, including an increased role for radiologists in disease prevention.

“I think the world is going to open up to how we can reduce women's risk with much greater precision than when we've been able to do before, and that's where it's really exciting to think of the imaging scientists and radiologists moving outside and beyond their really important role of detecting and diagnosing disease but actually playing a critical role in health promotion and disease prevention. That's where I get really excited,” emphasized Dr. Lehman.