AI-Driven Risk Assessment is Transforming Breast Cancer Screening

The integration of artificial intelligence (AI) into clinical practice is shifting the paradigm of preventative oncology from a one-size-fits-all approach to highly personalized care. In this interview, Elizabeth Mittendorf,MD, chief of Multi-Disciplinary Oncology, co-leader of the Parker Institute for Cancer Immunotherapy at Dana-Farber Cancer Institute, and the co-leader of the Breast Program for the Dana-Farber/Harvard Cancer Center, explored the clinical utility of Clarity Breast, an AI-driven technology that analyzes standard 2D mammograms to calculate a patient’s 5-year risk of developing breast cancer.

Unlike traditional methods, this tool operates independently of the radiologist’s standard interpretation, providing a second, objective data point to inform clinical decisions. Mittendorf detailed how this risk score can be used to triage screening backlogs by tailoring mammography intervals—potentially extending the time between screenings for those at lower risk while identifying candidates for enhanced surveillance via MRI.

Furthermore, Mittendorf addressed the critical need for equitable care, noting that the underlying technology for Clarity Breast has demonstrated accuracy across globally diverse populations, including racial and ethnic backgrounds where traditional models often underperform. From the potential of monitoring the efficacy of chemoprevention to the future state of subtype-specific risk prediction, Mittendorf highlighted the evolving role of AI in empowering both clinicians and patients with cancer.

CancerNetwork®: What is Clarity Breast?

Mittendorf: Clarity Breast is an innovative technology that uses artificial intelligence to look at a woman’s standard 2D mammogram, the mammogram they’d be getting for their routine screening, and it is able to calculate a woman’s 5-year risk of developing breast cancer.

How does Clarity Breast change that radiologist/surgeon consultation when predicting risk in a mammogram?

Clarity Breast, even though it’s AI applied to a mammogram, it makes you think that a radiologist would be involved. There doesn’t have to be a radiologist involved at all. In fact, as we’ve established [with] our program at the Beth Israel Deaconess Medical Center, the radiologist is interpreting that mammogram as they always have, independent of not knowing that Clarity risk score. These are 2 different pieces of information. First is the radiologist read, which, again, is the standard. This is what we do with screening mammograms. A report goes to the patient about that, and then a second, independent report that, again, the radiologist doesn’t have in their hands as they’re doing their initial interpretation, [is done] to suggest the 5-year risk.

This tool could help address the shortage of breast imagers in a busy clinical practice. How do you envision this risk score triggering the screening backlog?

That’s one of the things that I have the most enthusiasm for, and that is, how could we use it to address a shortage that we’re all experiencing nationwide for breast imagers? The current recommendation for most is to get an annual screening mammogram. You can see a future state where based on a person’s risk, determined by a number of their clinical features, their own personal history, their family history, other things, and this Clarity risk assessment, that we could triage patients for their screening based on the follow up, so maybe a woman who’s at a calculated low risk only gets a mammogram every other year, or maybe every 3 years.

There’s a lot of opportunity to think about this, and it’s fairly consistent with information that we got recently from the WISDOM trial. This was a study that was led by the team out at University of California San Francisco, Dr Laura Esserman, MD, MBS, that tried to ask the question, ‘Does everybody need an annual mammogram?’ The information that she had used––the patient’s clinical history, their family history, something called the polygenic risk score––and basically, we’re able to demonstrate that we can, in fact, tailor screening. Some people need more screening, so maybe not just an annual mammogram, but we need to do additional or enhanced screening, like MRI, and others can wait and have mammograms every other year. This fits very nicely into where the field is going, and perhaps gives a bit more objective data to inform those recommendations.

How should clinicians weigh the AI’s image-based score against established tools like the Gail or Tyrer-Cuzick models?

It’s just another part of the conversation. In our comprehensive breast center at Beth Israel Deaconess Medical Center, we have an outstanding team of physician assistants who do a rigorous risk assessment using a number of different questions that include things such as the Tyrer-Cuzick model, and that gives some information about a patient’s calculated risk. The Tyrer-Cuzick shows their 10 year risk and lifetime risk, and you can see where that might inform a recommendation to have the Clarity algorithm applied to their mammogram, or, let’s say they’re at an elevated risk, and then they get that Clarity score, and it suggests that this is enough data to inform a recommendation for chemotherapy prevention. Chemotherapy prevention is a pill that a woman might take every day to decrease her risk of developing breast cancer. It’s a pill that is effective, but it also has some adverse effects, so additional information as a patient considers the risks and benefits could help them come to the decision that feels right for them. It’s possible that, let’s say, a woman elects to take chemotherapy prevention, perhaps we could use this Clarity risk assessment to see if the prevention is working. Meaning, does that risk number go down?

Now…that’s something that I see as a future state, and certainly we would need to do the clinical studies to determine whether that is an efficacious approach to patients as they think about chemo prevention for their elevated risk.

For a patient with a normal mammogram but a high 5-year AI risk score, what is your current recommendation regarding supplemental?

At this point, I would view an elevated Clarity risk score, much like I view an elevated Tyrer-Cuzick risk, and I would have a conversation with the patient about enhanced surveillance to include what you described. It would probably be MRI in our practice, but an additional contrast-based exam. I would generally recommend that they be alternated, so they’d get a mammogram and MRI alternating every 6 months.

Traditional risk models often underperform in diverse populations due to a lack of representative data. Since this AI was trained on millions of images, have we seen data suggesting it provides a more equitable risk assessment across different racial and ethnic backgrounds?

There are available data around that. There was a manuscript published in the Journal of Clinical Oncology a couple years ago where the group that was based at Massachusetts Institute of Technology and Mass General Hospital developed the technology that effectively was leveraged to become the Clarity tool. It’s not the exact same tool, but it’s very comparable. What they did in that study was to look at a mammography-based risk score at number of institutions. There were sites in the US to include Massachusetts General, Novant, and Emory. Right there you can see there’s a different patient population. Just the percentage of Black women that are treated at Massachusetts General is very different than the percentage of Black women who are treated at Emory. It went a bit further. It also included international sites, one in Israel, one in Sweden, one in Taiwan and one in Brazil. All told, there was over 62,000 patients and over 128,000 mammograms. What they were able to demonstrate in that study was that this technology was accurate across this globally diverse population, and that the technology could offer broad and more equitable improvements in care for patients.

Do you see a path where this AI tool could eventually predict not just if a cancer will develop, but the specific subtype (e.g., Triple-Negative vs. ER+), allowing for even more targeted prevention?

Right now, when we suggest somebody is at increased risk, whether it be from a Tyrer-Cuzick or another assessment, the Clarity risk for chemotherapy prevention is endocrine therapy, and one could see where that might prevent the development of a hormone receptor–positive breast cancer. While that’s the most common type of breast cancer, accounting for approximately 70% of diagnoses, there’s other subtypes, triple-negative and HER2-positive. There’s a future state where, in fact, it could, and that’s because I believe it’s possible for the algorithms to continue to improve as they get more data. Now, to be clear as a tool, Clarity had to lock in their algorithm, so their algorithm is not continuously changing. It would have to be an academic group that would take that information and move forward with what you’re suggesting, which is to see if we could further refine the model to discriminate between the type of breast cancer that they’re at risk for. Now the flip side of that is we would then need to have strategies for prevention for those subtypes of disease. Here, I could see a future state where there are vaccines that target the HER2 protein for instance. If somebody would be at predicted risk for developing a HER2-positive breast cancer, maybe we would give them a vaccine. The vaccines are well tolerated, so that is an exciting future state, but the research would need to be done.

Reference

Eriksson M, Czene K, Vachon C, Conant EF, Hall P. Long-term performance of an image-based short-term risk model for breast cancer. J Clin Oncol. 2023;41(14):2536-2545. doi:10.1200/JCO.22.01564