Beyond Density: How AI Is Rewriting the Rules of Breast Cancer Screening
The modern practice of breast cancer screening was largely established by two cancer screening methods: the Papanicolaou stain for cervical cancer and mammography for early detection of breast cancer.
Of these two methods, mammography has had the most complicated and controversial history.
Because mammography involves a slight amount of radiation and also requires a follow-up biopsy, there has been extensive study of the optimal timing and frequency. Even today in 2026, there are major discrepancies between what the U.S. Preventive Services Task Force (an independent government body whose members are appointed by the Secretary of Health and Human Services) recommends in terms of biennial screening and the American College of Radiology, which recommends annual screening.
Artificial intelligence has now been formally recognized as having a major role in personalizing risk and screening in the April 2026 update of Breast Cancer Screening Guidelines of The National Comprehensive Cancer Network (NCCN), a nonprofit alliance of major U.S. cancer centers.
This is a transformative moment and Professor Regina Barzilay, the AI faculty lead at the MIT Jameel Clinic and affiliate at the MIT Computer Science and AI Laboratory, has been a leader in the science and AI modeling of breast cancer risk based on radiology image alone.
There are pre-existing methods of detecting increased risk that all suffer from disadvantages, most of which are overcome by the Mirai model, which was developed in collaboration with radiologists at Mass General Hospital.
One pre-existing method uses a patient’s menstruation duration, child bearing and breast feeding, family history, and other associated factors. But this process is time-consuming, expensive, and critically underestimates risk in Black women.
Much of the focus as of late has been on breast density, as seen on mammogram, which is hard to reproduce and has poor correlation with actual risk. But evidence supports additional screening in the smaller percentage of women with extremely dense breasts. Because half the U.S. population has dense breasts, it is not feasible to do additional screening on all women and as a result, there is a great deal of uncertainty amongst health care practitioners and women as well.
AI image modeling has now been recognized by NCCN as a legitimate risk factor in assessing level of risk in diverse populations. This provides a fantastic tool for allowing some women to know that they don’t have to return and face the cost, uncertainty, and stress of annual mammography. For others, it provides them with the knowledge to increase their vigilance with their annual mammograms.
In the U.S., the right to annual mammography is enshrined in federal law, but Mirai, which has been extensively validated internationally in about 2 million women, allows resource-constrained systems to develop schedules that maximize the value of mammography — catching those at higher risk while protecting the overall viability of the system.
Additional prospective randomized controlled trials are ongoing at two sites in the U.S. and two overseas sites (Israel and India) to see if the addition of breast MRI to the screening of women categorized as high-risk by Mirai can catch even more cancers than more frequent mammography.
For the extensive community of women and families touched by breast cancer and the fear of breast cancer, AI-derived risk scores can illuminate choices and give more personalized control of this critical health care process. The NCCN update represents formal recognition of this advance that can impact about 100,000 women who receive a mammogram each day once the model is implemented.
