June 24, 2026 | Researchers have found that breast cancer risk scores derived from mammograms change over time and differ distinctly between women who develop cancer and those who do not. The new research was published yesterday in Radiology (DOI: 10.1148/radiol.253023), a journal of the Radiological Society of North America (RSNA).
Most women diagnosed with breast cancer have no known genetic mutations or reported family history of the disease, so women and their doctors rely on routine mammogram screenings to flag changes to breast density that may be concerning.
Deep learning models are now used routinely to generate breast cancer risk scores from screening mammograms, using the entire image rather than a limited, predetermined feature such as density. AI image-based risk scores are incorporated into the 2026 National Comprehensive Cancer Network guidelines. For instance, the guidelines recommend that beginning at age 35, women with an elevated five-year risk score (greater than 1.7%) consider breast MRI in addition to annual mammography.
But researcher Constance D. Lehman MD, PhD, wondered what would be revealed if a woman’s five-year risk scores were tracked as she returned for regular mammograms. Lehman is a professor of radiology at Harvard Medical School and CEO of Clairity, Inc, which has an FDA-approved AI tool to give a five-year breast cancer risk score.
“Deep learning models have been primarily used to assess cancer risk scores at a static point in time,” she said in a press release about the work. “In this study, we evaluated longitudinal changes in the image-only deep learning breast cancer risk score using serial mammograms from a large screening cohort.”
Lehman and colleagues set up a retrospective study included women who underwent screening mammograms between 2009 and 2019 at six imaging sites spanning urban tertiary, community-based and rural practice settings all within the same healthcare system. All exams were standard 2D bilateral full-field digital mammography screening exams acquired with or without digital breast tomosynthesis on Hologic Selenia and Hologic Dimensions mammography platforms.
A total of 54,014 women (median age 61) were included in the dataset comprising 817 cancer patients and 53,197 cancer-free controls. Of the cancer patients, 451 (55%) were diagnosed with invasive cancer, 118 (14%) with ductal carcinoma in situ (DCIS), and 248 (30%) with an unknown cancer. Of these, 682 (83%) were mammogram-detected cancers, and 135 (17%) were detected between scheduled screenings. The researchers compared the risk scores of the 817 women diagnosed with cancer with the scores of the 53,197 cancer-free control individuals.
Each woman contributed one index exam (defined as the final screening mammogram within the year prior to a breast cancer diagnosis or the final mammogram in the five-year study period for cancer-free controls) and up to six prior annual mammograms, for a total of 158,807 mammograms. The median number of screening mammograms per woman was three.
The team used an image-only deep learning model, Mirai, to generate five-year breast cancer risk scores from each of the available mammograms. Mirai is an open-source research artificial intelligence (AI) model available at https://github.com/yala/Mirai. “In this research, only the model that is not approved by the U.S. Food and Drug Administration was used to analyze the four standard two-dimensional mammographic views using convolutional neural networks, and information was aggregated across views to produce a single risk estimate (0%–100%) per examination,” the authors explain in the paper. No demographic, clinical or historical imaging data were used.
The five-year risk scores for each individual mammogram were collected per patient, and the researchers used linear mixed effects models with random intercepts and slopes to characterize how these risk scores changed over time.
The researchers observed clinically relevant differences in risk trajectories between women who did and did not develop cancer. The slope of the curve of their risk scores changed as early as six years prior to diagnosis and became steeper over time. Among the cancer patients, AI risk scores increased progressively over the six years preceding diagnosis, with the median risk score for each mammogram increasing from 2.1 in the first five to six years of the study period to 6.6 at the index exam. Cancer-free women exhibited stable scores across all time points, with medians ranging from 1.8 to 2.2 over the study period.
“These findings demonstrate signals, invisible to the human eye, in the image alone can predict future risk,” Lehman said. “Our findings demonstrate that image-based AI risk scores evolve over time and that changes in those scores may provide additional information about future breast cancer risk,” she said.
Risk Prediction or Early Intervention?
The risk score trajectory among cancer patients was not strictly linear, though. It increased most steeply in the years immediately preceding the diagnosis. A gradual increase in scores in the early years of the study period was followed by a much steeper increase two years prior to the index exam.
In an editorial in the same issue of Radiology (DOI: 10.1148/radiol.261198), Ritse M. Mann, MD, PhD and Xin Wang, MSc ask if that finding suggests a change in risk or a change in reality.
“The most pronounced rise in breast cancer risk scores occurred in the years immediately preceding diagnosis, especially near the index examination,” they write. “This suggests that part of the apparent “risk” signal may be capturing early subclinical cancer rather than reflecting long-term susceptibility alone.”
That would change how patients are monitored, how screening intervals should evolve, and how we identify women who may benefit from supplemental imaging.
“Having a dynamic risk score opens up a whole new domain of more effective preventive therapies for breast cancer, similar to how we screen for and treat patients with high cholesterol and hypertension,” Lehman posited.
The work calls for a shift in how we think about risk, Mann and Wang agreed in their editorial.
“The importance of this study lies not only in confirming that deep learning models based on mammography can predict future breast cancer risk, but also in shifting attention from static, single-time-point risk assessment to dynamic, longitudinal risk assessment. Risk may no longer be viewed as only high or low. It may also be stable, slowly rising, or rapidly rising. These patterns may have different implications for screening intervals, supplemental imaging, and risk communication. In other words, the question may no longer be only ‘What is the individual’s current five-year risk?’ but also ‘How quickly is that risk changing relative to the individual’s own baseline?’”