She survived breast cancer. Now her AI tool could help you skip annual mammograms.

By Emily Bary

MIT computer-science professor Regina Barzilay lands on the MarketWatch 25 list of people who are reshaping how we live, work, save, spend and invest

As an MIT computer-science professor, Regina Barzilay was used to living on the bleeding edge of innovation, teaching computers to understand words in the nascent field of natural language processing. But when she was diagnosed with breast cancer in 2014, she was thrust into a different and, as she describes it, "really backwards" technological world.

That was the U.S. healthcare system, which Barzilay found unable to answer what she assumed would be basic data-analysis questions. She remembers asking her medical team how breast-cancer patients like her had fared at the hospital at which she was treated, but, even with so many medical records stored electronically, she got no good answers.

Barzilay went through chemotherapy, recovered and began to think about how her background in computer science could help other patients.

Now she sees that there was information in her first mammogram that could have tipped doctors off about her risk, even though her breast cancer didn't get diagnosed until the third annual scan. Barzilay knows this because she ran her first scan image through Mirai, an open-source artificial-intelligence tool she co-developed at MIT that better predicts cancer risk.

If Mirai and other tools like it are widely adopted, they could dramatically change the way we approach mammograms in this country. A recent study pegged the annual cost of breast-cancer screenings in the U.S. at $11 billion, but researchers like Barzilay see opportunities to bring costs down by tailoring screenings toward personalized understandings of risk, rather than one-size-fits-most guidelines. Barzilay lands on the MarketWatch 25 list of people who are reshaping how we live, work, save, spend and invest.

There's "a lot of waste going into screening everyone at the same rate," when "clearly we are not all at the same risk," Barzilay told MarketWatch. With the exception of patients who have certain genetic mutations, screening guidelines group women into "broad buckets" that focus on simple factors like age and breast density.

Mass adoption of AI screening tools could also help women avoid Barzilay's breast-cancer experience. Had she known early on what Mirai now reveals about her first scan, she thinks she could've been monitored more carefully, diagnosed earlier and perhaps spared the chemotherapy appointments that made her lose her hair.

"Today we really don't give a clear answer to women who are at risk, and I think the system really fails them," she said.

Enter Mirai

Barzilay's breast-cancer diagnosis got her interested in the ways AI could improve healthcare, but it wasn't a flawless start. She approached doctors about collaborations but found they struggled to understand how AI could be useful in their work. And her own initial research ideas, like studying the impact of nutrition on cancer, now strike her as "naive" because there was no readily available data for her to analyze.

Eventually Barzilay turned her attention to screening, where she saw a chance to develop smarter risk assessments. Before her diagnosis, Barzilay was told she was at higher risk for breast cancer because she had dense breasts, but since some 40% of American women have dense breasts but don't all develop breast cancer, that information wasn't so useful.

So far screening protocols have only improved dramatically for patients with the BRCA mutation, who have an increased risk of developing breast cancer. They are told to get monitored as early as 25. "It's still a very limited group, so you can really concentrate resources and help them, and it's totally changed the game in terms of survival for them," she said. For those without the mutation, however, she sees a need for better risk measures.

Even in the best-case scenarios, existing risk models are only 60% to 65% accurate, but Mirai's performance is closer to 80%, and without the same performance drops typically seen when looking at different subpopulations based on race or age.

Machines, in a way, "can see the future," according to Barzilay.

The Mirai model offers "a fundamental paradigm shift," said Hari Trivedi, a radiologist who codirects the Health Innovation and Translational Informatics machine-learning lab at Emory University. "It's much easier to get reliable and reproducible results using four images than it is to collect that data from patients who notoriously are forgetful and even between two years of screening may report their history differently."

Traditional risk models take into account a variety of factors, including a patient's age, breast density, family history of breast cancer, number of children, smoking history and breastfeeding history, before applying weights to all those measures and determining a five- to 10-year risk of developing cancer.

"These AI-based models don't use any of that information," said Trivedi. "The main difference here is that they are making their prediction based on the image alone." Mirai can see subtle changes in tissue, such as inflammation, that can't be easily detected by the human eye.

Trivedi got involved with Mirai when Barzilay and her team were looking for researchers to validate the model based on external data, since Mirai was trained on data from Massachusetts General Hospital that was dominated by white patients. AI models may perform well on the data they're trained on, but "really the proof of the value of it is, does it generalize to an external population?" he noted.

He found that Mirai did, outperforming traditional models while also generalizing better to minority patients.

There's a mix of "justified optimism and caution" around AI when it comes to healthcare professionals, said Diana Miglioretti, a biostatistician at the University of California, Davis. Speaking generally, she said it's "important that AI is rigorously evaluated in actual clinical practice to confirm real improvements in screening outcomes."

Right now, Mirai is a model that pharmaceutical companies and some labs have explored, according to Barzilay. She and her team chose to make Mirai publicly available because they wanted clinicians to understand how it worked.

"The goal is to really change clinical practice."

What change could look like

The most at-risk women, as indicated by the Mirai model, might benefit from being screened every six months instead of once a year. But by shifting toward more individualized assessments of risk, many more women might be fine undergoing screenings every two years.

Screening guidelines vary by health organization, but the American Cancer Society, for one, currently recommends annual mammograms for women between 45 and 54, with the option for annual screening beginning at 40 and a move to screening every two years once reaching 55.

"It's not just possible, but it's plausible, that this type of technology would eventually yield a screening strategy that is both more accurate and more cost-effective," Trivedi said. AI probably won't cut screenings in half, but even a reduction of 5% to 10% would have an impact by "sheer scale," given that some 40 million women get mammograms each year.

There would be psychological barriers to overcome, experts say. Some patients fail to comply with annual mammogram recommendations and might be relieved if guidelines were relaxed. Other women, though, might be nervous to be screened less regularly than is currently suggested, even if AI predicts they're at low risk.

"It actually makes [women] very uncomfortable to be told to screen less, and so it remains to be seen whether or not people would actually adhere to being told to screen less," said Kimberly Badal, a computational biologist at the University of California, San Francisco, whose research focuses on breast-cancer screenings.

Badal is "excited" about AI mammography models, and she's studying risk-based screening through her own work. But she's also cognizant of the fact that AI tools could widen health disparities, since the models ultimately could run on top of medical-record systems. That means accessing the tool would require "a lot of infrastructure" and "a lot of money."

"Community healthcare hospitals are not going to have access to it," she said, citing the costs and complexities. "The majority of the world actually is not going to have access to it."

Adoption even at high-tech medical centers could be slow, particularly because of the conflicting financial incentives in the U.S. fee-for-service healthcare model. Insurers might be willing to pay for something that reduces the number of screenings, but hospitals and radiologists might be less motivated, Trivedi noted.

In general, he thinks healthcare professionals are frustrated by the bureaucracy required to get AI systems implemented. "At least half of the reason that AI is not being adopted as fast as it could be" is that the AI models simply aren't good enough, he said. But the other big reason is the "sludge" in the U.S. healthcare system, including questions about who pays and how technologies are integrated.

There are signs that AI could become a more prevalent screening tool in the coming years. Clairity, an AI platform that also uses mammogram imagery to predict breast-cancer risk going out five years, has gotten de novo authorization from the Food and Drug Administration, which is "a piece of tangible evidence" that the FDA considers this sort of risk prediction to be "safe and valid," according to Trivedi.

Beyond breast cancer

Though Barzilay has a personal connection to breast-cancer research, she's become an ambassador for the adoption of AI in medicine more broadly.

(MORE TO FOLLOW) Dow Jones Newswires

11-11-25 0800ET