Related Experiment Video
Updated: Aug 13, 2026
06:46
Competing-Risk Nomogram for Predicting Cancer-Specific Survival in Multiple Primary Colorectal Cancer Patients after Surgery
Published on: September 27, 2024
Predicting 5-Year Breast Cancer Risk From Longitudinal Digital Breast Tomosynthesis: A Single-Center Retrospective
Yanqi Xu1, Laura Heacock2, Jungkyu Park3
1Center for Data Science, New York University.
AJR. American Journal of Roentgenology
|August 12, 2026
Summary
A new deep learning model using longitudinal digital breast tomosynthesis (DBT) images significantly improves breast cancer risk prediction compared to older methods. This advancement offers potential for more personalized breast cancer screening strategies.
Area of Science:
- Radiology
- Artificial Intelligence
- Oncology
Background:
- Traditional breast cancer risk models rely on full-field digital mammography (FFDM).
- Digital breast tomosynthesis (DBT) is now a common screening tool, but its use in long-term risk prediction is understudied.
- This research addresses the gap in understanding DBT's potential for predicting future breast cancer risk.
Purpose of the Study:
- To develop and validate a deep learning model for predicting long-term breast cancer risk using longitudinal DBT data.
- To assess the model's performance against existing methods, including single-timepoint DBT, FFDM-based models, and clinical risk assessment tools.
Main Methods:
- A retrospective analysis of 313,335 DBT examinations from 161,077 women was conducted.
- A deep learning model (DRP) was created to estimate 2- to 5-year breast cancer risk, incorporating longitudinal DBT images, patient age, and breast density.
- Model performance was evaluated using AUC, time-dependent concordance index, and integrated Brier score, comparing it with single-timepoint DBT, Mirai (FFDM-based), and Tyrer-Cuzick models.
Main Results:
- The longitudinal DRP model demonstrated superior 5-year risk prediction (AUC = 0.721) compared to single-timepoint DBT (AUC = 0.707) and the Mirai model (AUC = 0.687).
- In a case-control cohort, the DRP model (AUC = 0.676) outperformed the Tyrer-Cuzick model (AUC = 0.563).
- The model showed nuanced risk stratification in women with dense or fatty breasts, correlating with observed cancer incidence.
Conclusions:
- Deep learning models utilizing longitudinal DBT data enhance long-term breast cancer risk prediction beyond current FFDM-based and clinical models.
- Longitudinal DBT-based risk prediction can facilitate dynamic risk assessments and personalized screening approaches.
- This approach holds promise for improving early detection and management of breast cancer.