Related Experiment Video
Updated: Aug 5, 2026
15:48
Tracking the Mammary Architectural Features and Detecting Breast Cancer with Magnetic Resonance Diffusion Tensor Imaging
Published on: December 15, 2014
DBT-DINO: Toward Foundation Model-Based Analysis of Digital Breast Tomosynthesis
Felix J Dorfner1,2, Manon A Dorster1, Ryan Connolly3
1Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School, 149 Thirteenth St, Charlestown, MA 02129.
Radiology
|August 4, 2026
Summary
A new foundation model, DBT-DINO, shows strong performance in classifying breast density from digital breast tomosynthesis images. However, domain-specific pretraining did not significantly improve 5-year breast cancer risk prediction or lesion detection compared to a general baseline.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Foundation Models for 3D Modalities
- Digital Breast Tomosynthesis (DBT) Analysis
Background:
- Foundation models show promise in medical imaging but are underexplored in 3D modalities.
- Digital breast tomosynthesis (DBT) is widely used in breast cancer screening, yet lacks a dedicated foundation model.
- Existing models often use general image datasets (e.g., ImageNet), potentially limiting performance on specialized medical data.
Purpose of the Study:
- To develop and evaluate DBT-DINO, a foundation model specifically for DBT images.
- To assess the impact of domain-specific pretraining on DBT data for various clinical tasks.
- To compare DBT-DINO's performance against a general ImageNet-pretrained baseline.
Main Methods:
- Retrospective study using DBT images from Mass General Brigham (2011-2024).
- Self-supervised pretraining of the DINOv2 methodology on over 25 million DBT sections from ~28,000 patients.
- Evaluation on three downstream tasks: breast density classification, 5-year breast cancer risk prediction, and lesion detection.
Main Results:
- DBT-DINO achieved 79% accuracy in breast density classification, outperforming the DINOv2 baseline (73%, P < .001).
- No significant difference was observed in 5-year breast cancer risk prediction (AUC 0.78 vs 0.76, P = .057).
- No significant difference was found in lesion detection sensitivity (62% vs 67%, P = .60).
Conclusions:
- DBT-DINO demonstrates significant potential for breast density classification using DBT images.
- Domain-specific pretraining did not yield superior results for risk prediction or lesion detection tasks.
- Further refinement of domain-specific pretraining may be necessary for localized detection tasks in DBT analysis.
