A deep-learning (DL) model may help estimate risk of future breast cancer, especially for women with dense breasts, according to results published May 4 in JAMA Network Open.
The model outperformed breast density in estimating future risk and stratified false-negative screening results across risk groups, wrote a team led by Leslie Lamb, MD, of Massachusetts General Hospital in Boston.
“Findings of this study suggest that DL risk models could offer a more precise and equitable alternative to breast density as a policy criterion for determining access to supplemental breast imaging,” Lamb and colleagues wrote.
With imaging facilities in the U.S. now required to disclose information to women on their breast density, states have introduced and passed laws to expand insurance coverage for women who need supplemental imaging along with their regular mammogram.
The researchers highlighted some concerns with this approach, such as overuse of imaging resources, increased false-positive results, and additional financial burdens. They called for more precise risk stratification tools that go beyond breast density alone for more personalized and fair screening.
The Lamb team compared the performance of a DL breast cancer risk model (Mirai, Massachusetts Institute of Technology and Massachusetts General Hospital) to that of breast density evaluated by radiologists to estimate future breast cancer and false-negative results.
The multisite study included over 123,000 mammograms in over 67,000 women with a median age of 58 years. Of the total mammograms, about 51,000 were classified as dense. The women underwent mammography screening between 2009 and 2018 and follow-up exams through 2023.
The model achieved a higher area under the receiver operating curve (AUROC) over breast density (0.71 versus 0.53; p < 0.001). It also led to significantly higher false-negative rates increased across DL risk groups. This included 2.1 per 1,000 exams in high-risk women versus one and 0.6 in intermediate- and low-risk groups, respectively.
Furthermore, women with dense breasts had higher false-negative rates than women with non-dense breasts (1.7 versus 0.6 per 1,000 exams; p < 0.001), the researchers found.
Finally, the team reported that the model’s performance did not improve when breast density was added to it.
The results support precision risk-based screening strategies using DL models, the study authors wrote.
“In the current era of density legislation, DL models could also be used to identify the subset of women with dense breasts most likely to benefit from supplemental imaging while sparing low-risk women unnecessary imaging,” they added.
The authors also called for future research to focus on continuous model development, prospective implementation of DL risk assessment in breast cancer screening, and evaluation of how this can affect clinical outcomes, cost-effectiveness, and patient experience.
Read the full study here.
