Deep Learning Model Accurately Predicts 5-Year Breast Cancer Risk Using 3D Mammography
Researchers have developed and validated a deep learning model utilizing longitudinal 3D digital breast tomosynthesis examinations to predict individualized five-year breast cancer risk, offering a significant methodological advance for early screening protocols.
- A retrospective study validated an artificial intelligence model using 29,946 patient records collected between January 1, 2004, and December 31, 2020, at the Netherlands Cancer Institute.
- The refined analytical dataset comprises 9,133 patients—including 2,562 biopsy-proven breast cancer cases within ten years and 6,571 intermediate-risk patients with at least ten years of negative screening follow-up.
- By integrating consecutive longitudinal tomosynthesis scans with structured electronic medical record data such as BI-RADS scores and breast density ACR categories, the algorithm successfully evaluates short- and long-term malignancy probabilities.
Longitudinal Tomosynthesis Dataset and Validation Parameters
Investigators structured the evaluation in line with TRIPOD-AI recommendations for transparent reporting of multivariable prediction models. The primary dataset captured consecutive digital screening mammograms from the hospital archive. Exclusion criteria omitted patients lacking at least one year of screening follow-up.
The final patient cohort of 9,133 individuals was randomly partitioned at the patient level into training, validation, and test subsets with a 7.5 to 1 to 1.5 ratio. This split ensured that images from any single patient remained exclusive to one partition. Specifically, the training, validation, and test sets incorporated 6,858, 919, and 1,356 patients, corresponding to 32,049, 4,432, and 6,311 individual examinations, respectively. For patients with prior breast disease, the model incorporated longitudinal pathology and therapy reports, tracking pathologic tumor stage, node stage, and hormone receptor status including estrogen and progesterone receptor profiles.
Integration of Clinical Risk Factors and Imaging Biomarkers
The deep learning architecture moves beyond static two-dimensional risk calculators by processing sequential volumetric images alongside structured clinical variables. Self-reported risk metrics including age, race, family history, menopausal status, and age of menarche were compiled alongside genetic determinants and prior histories of ovarian or breast malignancies. Radiologists assigned BI-RADS final assessment scores—ranging from normal to highly suggestive of malignancy—and categorized breast density using American College of Radiology grades, stretching from fatty tissue to extremely dense parenchyma. Missing density data points were resolved via nearest neighbor interpolation using adjacent screening images.
Methodological Implications for Screening Protocols
Traditional risk models often rely solely on tabular epidemiological questionnaires, which frequently lack the nuanced architectural detail captured by volumetric imaging. By evaluating consecutive 3D examinations, the deep learning model detects subtle parenchymal changes over time that precede clinically detectable masses. This longitudinal tracking supports more precise interval planning for high-risk cohorts.
Future clinical deployment will require prospective validation across diverse healthcare systems to ensure generalizability beyond single-institution cohorts.
*Disclaimer: The information provided in this article is for educational and scientific communication purposes only and does not constitute medical advice. Always consult with a qualified healthcare provider regarding any medical condition, diagnosis, or treatment plan.*