New research demonstrates that image-only deep learning models can identify breast cancer risk by tracking progressive changes in mammogram features over time.
Researchers have identified that image-based risk scores derived from screening mammograms exhibit distinct longitudinal trajectories that differentiate patients who develop breast cancer from those who remain cancer-free. This study, published in the journal Radiology, utilizes deep learning architectures to extract predictive signals directly from raw imaging data, bypassing the limitations of traditional, static risk assessment models that rely heavily on breast density or clinical history.
The investigation analyzed 158,807 screening mammograms collected between 2009 and 2019 from a diverse cohort of 54,014 women. Constance D. Lehman, M.D., Ph.D., professor of radiology at Harvard Medical School and CEO of Clairity, Inc., spearheaded the research, which leveraged a validated, open-source image-only deep learning model to generate continuous five-year risk scores. The methodology intentionally excluded demographic, clinical, and historical imaging data to isolate the predictive power inherent in the pixel-level information of the mammograms themselves.
The study cohort included 817 patients diagnosed with invasive cancer or ductal carcinoma in situ within one year of their index exam, alongside 53,197 cancer-free controls. By applying the deep learning model to serial mammograms, the team observed that risk scores for cancer patients began to diverge from the baseline as early as six years prior to clinical diagnosis. While cancer-free individuals maintained stable, flat risk trajectories with median scores ranging from 1.8 to 2.2, the cancer cohort showed a progressive, statistically significant increase in scores over the same period.
The data revealed a non-linear acceleration in risk scores among the cancer group, particularly in the two years immediately preceding a diagnosis. Median risk scores for these patients rose from 2.1 in the early stages of the study to 6.6 at the time of the index exam. This trend remained consistent across various patient subgroups, including those categorized by different age ranges and breast density profiles, suggesting high generalizability for the model across diverse clinical settings.
These findings indicate that deep learning models can detect subtle, non-visual features within standard 2D bilateral full-field digital mammography that precede the development of detectable lesions. By focusing on the entire image rather than predetermined features, the model captures complex patterns that traditional risk models often overlook. This approach effectively addresses the limitations of current population-based screening, where most patients diagnosed with breast cancer lack known genetic mutations or significant family histories.
The integration of these dynamic risk scores into clinical workflows represents a shift toward personalized, risk-based screening strategies. Dr. Lehman emphasized that this methodology mirrors established clinical practices for managing chronic conditions like hypertension or hypercholesterolemia. By treating imaging data as a dynamic biomarker, clinicians can potentially implement more aggressive surveillance or preventive interventions for patients whose longitudinal trajectories indicate an elevated risk profile.
Deep learning models have been primarily used to assess cancer risk scores at a static point in time. 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.
The significance of this research lies in its potential to mitigate existing disparities in screening performance by reducing reliance on inconsistent self-reported data. Because the model relies exclusively on imaging, it provides a standardized, objective metric that is less susceptible to the biases inherent in clinical questionnaires. This capability is particularly vital for patients in rural or community-based settings where access to comprehensive genetic testing or detailed clinical records may be limited.
The 2026 National Comprehensive Cancer Network guidelines have already begun incorporating these AI-derived risk scores into their clinical recommendations. Current protocols suggest that women aged 35 and older with a five-year risk score exceeding 1.7% should consider supplemental breast MRI alongside annual mammography. As these models continue to be deployed across U.S. healthcare institutions, the focus will shift toward validating the clinical outcomes of these risk-stratified screening pathways.
Future work will likely focus on refining the model’s sensitivity to even earlier indicators of malignancy and integrating these scores into broader multi-modal diagnostic frameworks. The ability to extract predictive data from standard imaging archives suggests that existing hospital databases could be retroactively analyzed to improve early detection rates. Researchers are now monitoring how these longitudinal scores influence long-term patient outcomes and the efficacy of preventive therapies in high-risk populations.
