Artificial intelligence risk scores derived from screening mammograms may change over time and follow different patterns in women who later develop breast cancer. Research published in Radiology assessed serial mammograms from women screened at six sites within a large academic health system. Image-only deep learning scores increased progressively among women diagnosed with invasive cancer or ductal carcinoma in situ, while remaining broadly stable among those who stayed cancer-free. The difference was visible several years before diagnosis and became most pronounced close to the final screening examination, supporting further assessment of these scores as dynamic imaging biomarkers rather than fixed, single-time-point estimates.
Serial Mammograms Track Changing Risk
The retrospective cohort included women who underwent screening mammography between 2009 and 2019 across urban tertiary, community and rural settings. The final dataset contained 158,807 mammograms from 54,014 women, with a median age of 61 years. Within one year of their final included screening examination, 817 women were diagnosed with invasive breast cancer or ductal carcinoma in situ. The remaining women had no breast cancer diagnosis during follow-up.
Each woman contributed an index mammogram and up to six prior annual examinations. The index examination was the final screening mammogram within the year before diagnosis for women with cancer, or the examination marking the end of five years of follow-up for cancer-free controls. One mammogram was retained for each yearly interval to provide a consistent longitudinal structure.
A validated image-only deep learning model generated a continuous five-year breast cancer risk estimate from four standard mammographic views. It used image pixels from the current examination without demographic, clinical or historical imaging data. Linear mixed-effects models accounted for repeated examinations and differences in follow-up timing. The primary comparison assessed whether annual score trajectories differed between women who developed cancer and those who remained cancer-free. Additional comparisons examined patterns by age and breast density.
Women who developed cancer were older and more often had dense breasts, a personal or family history of breast cancer and postmenopausal status. Age-adjusted modelling tested whether age differences explained the contrasting trajectories.
Risk Scores Diverge Before Diagnosis
Among women who developed breast cancer, median risk scores rose from 2.1 five to six years before the index examination to 6.6 at the index examination. Scores among cancer-free women remained comparatively stable, ranging from 1.8 to 2.2 across the same period. Women later diagnosed with cancer also had higher scores at every yearly interval examined.
The strongest within-person increases appeared closest to diagnosis. Scores changed little across more distant annual intervals but rose more clearly during the final two years. From one year before the index examination to the index examination, the median change was 1.4 among women with cancer and zero among cancer-free controls. The cancer trajectory therefore became progressively steeper as diagnosis approached, while the control trajectory remained essentially flat.
Must Read: Mammographic Density and Inherited Breast Cancer Risk
Longitudinal modelling confirmed the separation. Risk scores increased by an estimated 1.13 units per year among women who developed cancer, compared with 0.09 units among cancer-free women. The difference between the annual slopes was 1.04 units and remained nearly unchanged after adjustment for age.
The pattern was consistent across age categories and among women with dense and nondense breasts. Annual score increases were greater in the cancer group in every subgroup examined, so the divergence was not confined to a particular age range or breast density category within the cohort.
Dynamic Scores Could Support Risk-Adapted Care
Traditional breast cancer risk tools rely on genetic, family and reproductive information that is relatively fixed. Mammographic density can change over time. Image-only deep learning models use the full mammographic image, allowing serial examinations to generate updated risk estimates.
The rising scores before diagnosis suggest that mammographic patterns linked to future cancer risk may evolve years before malignancy becomes radiographically apparent. Serial assessment could provide information beyond a single five-year estimate. Potential uses include personalised screening intervals, supplemental imaging decisions and preventive strategies. Prospective evaluation is required before longitudinal changes can be incorporated into clinical decision-making.
Applying a single-examination model at each screening visit may avoid dependence on complete longitudinal image sequences. Models requiring prior mammograms as direct inputs may be harder to implement where screening intervals vary or earlier records are incomplete. Repeated use of a validated point-in-time model offers a potentially more scalable way to observe changes relative to each woman’s previous score.
Several limitations constrain the findings. The cohort came from one healthcare system and used mammography equipment from a single manufacturer, despite covering tertiary, community and rural sites. Additional work is needed to confirm whether the temporal patterns generalise to other populations, systems and equipment. Some women may also have contributed remote examinations to the model’s original training dataset. The statistical approach estimated linear trajectories, while future work may assess nonlinear models. The analysis did not test diagnostic detection or establish thresholds for action.
Image-based deep learning risk scores followed different paths in women who did and did not develop breast cancer. Scores remained stable in cancer-free women but rose progressively before diagnosis, with the steepest increases occurring close to the final screening examination. The pattern persisted across age and breast density groups and was not explained by age differences. These findings support further evaluation of serial artificial intelligence risk scores as dynamic biomarkers for personalised screening and risk reduction. Prospective work is needed to determine how score changes should inform clinical decisions and whether the results extend beyond the health system and equipment represented.
Source: Radiology
Image Credit: iStock
References:
Lehman CD, Mercaldo SF, Azam S et al. (2026) Longitudinal Analysis of Changes in Deep Learning Image-based Breast Cancer Risk Scores over Time. Radiology; 319:3.
