Breast cancer risk prediction has long relied on static snapshots — a single mammogram read at a single point in time. A model that instead watches how subtle asymmetries between breasts evolve across sequential screenings could fundamentally change who gets flagged for early intervention, and how far in advance.
Researchers constructed STA-Risk, a deep learning architecture designed to encode three distinct signals simultaneously: left-right breast asymmetry at any given exam, changes in that asymmetry across time, and a customized loss function that penalizes the model for ignoring those asymmetric cues. The system was trained and validated on sequential mammograms from the CSAW-CC dataset — 406 cancer cases and over 6,000 controls — and then stress-tested on a fully independent cohort of 293 cancer cases and 297 controls. STA-Risk achieved concordance indexes of 0.72 and 0.73 on the two datasets, respectively, outperforming all benchmark models tested, which ranged from 0.67 to 0.72. Critically, the advantage held across all prediction windows from one to five years out, and ablation studies confirmed that each of the three architectural components contributed meaningfully to performance.
The broader significance here is methodological. Most clinical risk tools — Tyrer-Cuzick, BCSC models, even prior deep learning approaches — treat mammograms as independent images. STA-Risk treats screening as a longitudinal process, which more closely mirrors how radiologists actually think when they compare current films to priors. The concordance index gains, while modest in absolute terms, are clinically relevant in a population-screening context where even small improvements in discrimination translate to thousands of women correctly reclassified. Key limitations include restriction to Hologic imaging systems, a retrospective case-control design that may not reflect true screening population distributions, and the need for at least two prior exams per patient. Validation across diverse imaging hardware and prospective cohorts will be essential before this approach influences clinical triage. Directionally, this work is a credible step toward dynamic, individualized breast cancer surveillance.