DBT-DINO: Toward Foundation Model-Based Analysis of Digital Breast Tomosynthesis.
Authors
Affiliations (7)
Affiliations (7)
- Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School, 149 Thirteenth St, Charlestown, MA 02129.
- Department of Radiology, Charité - Universitätsmedizin Berlin corporate member of Freie Universität Berlin and Humboldt Universität zu Berlin, Berlin, Germany.
- Mass General Brigham Data Science Office, Boston, Mass.
- Department of Computer Science, Institute for Machine Learning, ETH Zürich, Zürich, Switzerland.
- Massachusetts General Hospital Cancer Center and Harvard Medical School, Boston, Mass.
- Department of Radiology, Massachusetts General Hospital, Boston, Mass.
- Department of Radiology, University of Wisconsin School of Medicine and Public Health, Madison, Wis.
Abstract
Background Foundation models show promise in medical imaging but remain underexplored in three-dimensional modalities. Despite the adoption of digital breast tomosynthesis (DBT) in breast cancer screening, no dedicated foundation model currently exists for this modality. Purpose To develop and evaluate a foundation model for DBT (DBT-DINO) and assess the impact of domain-specific pretraining across multiple clinical tasks. Materials and Methods This retrospective study used DBT images from Mass General Brigham acquired between March 2011 and February 2024. Self-supervised pretraining was performed using Meta AI's DINOv2 methodology on more than 25 million two-dimensional sections from 487 975 DBT volumes from 27 990 patients. Three downstream tasks were evaluated: <i>(a)</i> breast density classification using 5000 screening examinations, <i>(b)</i> 5-year risk of developing biopsy-proven breast cancer using 106 417 screening examinations, and <i>(c)</i> lesion detection using 393 annotated volumes. The performance of DBT-DINO was compared with that of ImageNet-pretrained DINOv2 baselines using McNemar and DeLong tests. Results A total of 4981 patients (mean age, 57.76 years ± 11.40 [SD]; 4855 female) were included for density classification, 31 561 patients (mean age, 60.09 years ± 10.47; 31 559 female) were included for risk prediction, and 199 female patients were included for lesion detection. For breast density classification, DBT-DINO achieved 79% (786 of 997 examinations) accuracy, outperforming the DINOv2 baseline (73% [728 of 997 examinations]; <i>P</i> < .001). For 5-year breast cancer risk prediction, DBT-DINO had an area under the receiver operating characteristic curve (AUC) of 0.78 and DINOv2 had an AUC of 0.76 (<i>P</i> = .057), showing no evidence of a difference. In lesion detection, DINOv2 had an average sensitivity of 67% (91 of 136 lesions), whereas DBT-DINO had a sensitivity of 62% (84 of 136 lesions) (<i>P</i> = .60), again with no evidence of a difference. Conclusion DBT-DINO demonstrated strong performance in breast density classification; however, there was no evidence of a difference compared with the ImageNet baseline in 5-year breast cancer risk prediction or lesion detection, suggesting that domain-specific pretraining for localized detection tasks required further refinement. © RSNA, 2026 <i>Supplemental material is available for this article.</i> See also the editorial by Wu in this issue.