Continuing to affirm the potential merits of AI in risk stratification for breast cancer, new research suggests that AI may predict the development of breast cancer up to 10 years prior to diagnosis.
For the retrospective study, recently published in Radiology, reviewed mammography data for 31,394 women (median age of 57.6) who had a total of 88,963 mammography exams. Mammograms for up to 10 years prior to diagnosis were included for all women, according to the study. The study authors subsequently evaluated the use of three AI algorithms for early detection of breast cancer, including Vara version 2.1 (Vara AI), Lunit Insight MMG version 1.1.7.2 (Lunit) and MammoScreen version 2.1.0 (Therapixel).1
At a 90 percent specificity, the researchers found that the AI software modalities flagged between 12.7 to 17 percent of potential cancers 10 years prior to diagnosis.1
Six years prior to diagnosis, the AI algorithms detected between 19 and 19.7 percent of potential breast cancer, according to the study authors. The AI modalities continued to improve, detecting between 24.2 and 25.2 percent of potential breast cancer four years prior to diagnosis.1
“All three systems demonstrated a consistent increase in per-examination AI scores for individuals with breast cancer over the 10 years leading up to diagnosis, while AI scores for cancer-free individuals remained stable over that time,” noted lead study author Sarah Hickman, MD, PhD, who is affiliated with Barts Health NHS Trust in London, United Kingdom, and the Department of Oncology and Pathology at Karolinska Institute in Stockholm, Sweden, and colleagues.
The study authors also compared the early detection of the AI algorithms with the predictive capacity of breast density.
At five to six years prior to diagnosis, the AI software modalities offered an AUC ranging between 60 to 64 percent in comparison to 57 percent for an AI density score. At three to four years prior to diagnosis, the researchers noted an AI software algorithm AUC ranging between 67 to 70 percent while the AUC for the AI density score was 56 percent. The study authors pointed out up to a 14 percent difference in AUC for the AI algorithms (ranging from 72 to 76 percent) one to two years prior to diagnosis in comparison to the AI density score (62 percent).1
“For each time interval from 5–6 years before diagnosis onwards, the AUC for any AI score was higher than the AUC for mammographic density,” observed Hickman and colleagues.
Three Key Takeaways
• AI demonstrates long-range predictive value for breast cancer. Three separate AI algorithms all showed a consistent, progressive increase in risk scores on mammograms up to 10 years before a breast cancer diagnosis while scores for cancer-free patients remained stable. This suggests AI may provide a meaningful early warning signal well before conventional detection.
• AI outperforms breast density as a risk stratification tool. Across all evaluated time intervals from five to six years pre-diagnosis onward, the AI algorithms achieved higher AUC values than AI-derived density scores. The gap widened closer to diagnosis, with AI algorithms reaching AUCs of 72–76 percent compared to just 62 percent for density scores one to two years prior, a difference of up to 14 percent.
• AI may help guide supplemental imaging decisions. The authors suggest that AI CAD scores could be combined with traditional mammographic density measurements to better determine which patients need supplemental imaging (e.g., MRI). This has practical clinical implications for moving beyond density-alone criteria, though limitations, including an enriched dataset and lack of cancer localization correlation, may temper implementation.
While acknowledging the established use of breast density as a “filtering mechanism” for supplemental imaging, the authors suggested increased utilization of AI for determining the necessity of supplemental imaging.1
“One suggestion is to combine the AI CAD score with traditional mammographic density to account for risk and masking. Another suggestion is to train additional deep learning systems to explicitly assess long-term risk and masking potential with strategies similar to those used in AISmartDensity, which achieved an incremental cancer detection rate of 64 per 1000 MRI examinations.2
(Editor’s note: For related content, see “Mammography Study Shows Impact of AI-Powered Slab Reconstruction with DBT,” “Can Adjunctive AI Enhance DBT Detection of Invasive Lobular Cancer?” and “Video: Wendie Berg, MD, PhD, Discusses the Recent ACP Guidance on Breast Cancer Screening.”)
In regard to study limitations, the authors acknowledged use of an enriched dataset and a lack of evaluation for correlation between AI scores and localization of cancer.
References
- Hickman S, Gialias P, Cossio F, et al. Artificial intelligence detection scores in screening mammography for early breast cancer alerts. Radiology. 2026;319(3). https://doi.org/10.1148/radiol.251309 .
- Liu Y, Sorkhei M, Dembrower K, Azizpour H, Strand F, Smith K. Use of an AI score combining cancer signs, masking, and risk to select patients for supplemental breast cancer screening. Radiology. 2024;311(1):e232535. doi: 10.1148/radiol.232535.
