Longitudinal DBT + AI Improves Breast Cancer Risk Prediction, Suggests Study
USA: Researchers have found in a new study that a deep-learning model using longitudinal digital breast tomosynthesis (DBT) exams may improve prediction of 2–5-year breast cancer risk compared with conventional mammography-based and clinical risk models.
In 161,077 women (313,335 DBT exams), the DBT model incorporated serial imaging, age, and breast density. For 5-year risk prediction, the DBT model achieved an AUC of 0.72, outperforming single-timepoint DBT (0.71), Mirai using digital mammography (0.69), and Tyrer-Cuzick (0.56 in the matched cohort).
Serial DBT may capture changes in breast density and tissue texture over time, providing risk information beyond a single mammogram. The model also showed potential for risk stratification in women with extremely dense or fatty breasts.
Longitudinal DBT combined with AI could enable dynamic, personalized breast cancer risk assessment and potentially help tailor future screening strategies. However, external validation across diverse populations, institutions, and imaging systems is needed before clinical implementation.
The findings were reported in the American Journal of Roentgenology by Yanqi Xu, MSc, Center for Data Science, New York University, and colleagues. The researchers noted that although imaging-based breast cancer risk prediction has traditionally relied on full-field digital mammography (FFDM), DBT is now widely used for breast screening in the United States. Its ability to predict cancer risk several years into the future, however, has been less extensively studied.
For the retrospective study, investigators analyzed DBT examinations performed between January 2016 and August 2020 at a single health institution. The participants had a mean age of 58.5 years. Researchers developed the DBT-based risk prediction (DRP) model using serial DBT examinations along with patient age and breast density to estimate the likelihood of developing breast cancer within 2 to 5 years.
The study led to the following findings:
- In 34,570 women, the longitudinal DRP model achieved a 5-year AUC of 0.721, outperforming single-timepoint DBT (0.707) and the Mirai model using digital mammography (0.687).
- In a matched cohort of 432 women, the longitudinal DBT model had a 5-year AUC of 0.676 versus 0.563 for the Tyrer-Cuzick model.
- Among women with extremely dense breasts, 39.7% were classified as average risk, with a 5-year breast cancer incidence of 0.8%.
- Among women with fatty breasts, 14.8% were classified as high risk, with a 5-year breast cancer incidence of 2.6%.
The authors suggested that repeated DBT examinations could provide additional information by capturing changes in breast density and breast tissue characteristics over time. Such longitudinal imaging data may complement established clinical risk factors and support more individualized approaches to screening.
However, the study was conducted at a single institution, highlighting the need for validation in broader, more diverse populations and across different imaging platforms before the model can be incorporated into routine clinical practice.
Reference:
https://ajronline.org/doi/10.2214/AJR.26.34951
Next Story