An AI model based on digital breast tomosynthesis (DBT) images performed strongly in classifying breast density, according to research published August 4 in Radiology.
The model, which used the second version of Meta’s self-DIstillation with NO labels (DINO) algorithm, accurately classified about four of out five DBT exams by density. Researchers led by Christopher Bridge, DPhil, from Massachusetts General Hospital and Harvard Medical School in Boston, named the model DBT-DINO.
“These results indicate that in-domain pretraining and self-supervised learning techniques have substantial potential for analyzing DBT images, but pretraining approaches for localized detection tasks require further refinement,” Bridge and colleagues wrote.
Foundation models in recent years have shown promise in medical imaging. However, the researchers noted a lack of exploration toward 3D imaging modalities. This includes DBT, which has grown in adoption since its first regulatory approval for clinical use in the U.S. in 2011.
The Bridge team developed its own foundational model to address this knowledge gap with DBT-DINO. The team used DINOv2’s self-supervised pretraining on more than 25 million 2D sections from 487,975 DBT volumes from 27,990 women.
The researchers evaluated three downstream tasks for the model: breast density classification using 5,000 screening exams; five-year risk of developing biopsy-proven breast cancer using 106,417 screening exams; and lesion detection using 393 annotated volumes.
Finally, the team compared the performance of DBT-DINO with that of ImageNet-pretrained DINOv2 baselines.
The study included 4,981 women for density classification, 31,561 women for risk prediction, and 199 women for lesion detection. It also included DBT images collected between 2011 and 2024.
While DBT-DINO outperformed ImageNet-pretrained DINOv2 for classifying breast density, both models had comparable performances for cancer risk prediction and lesion detection.
|
Performances of foundation models on DBT exams |
|||
|
Measure |
ImageNet-pretrained DINOv2 |
DBT-DINO |
P value |
|
Breast density classification |
73% |
79% |
|
|
Five-year breast cancer risk prediction (area under the curve [AUC]) |
0.76 |
0.78 |
0.057 |
|
Lesion detection |
62% |
67% |
0.6 |
“Although we can achieve performance improvements on global tasks such as breast density prediction and risk prediction, performance in very localized tasks such as lesion detection did not benefit from our in-domain continued pretraining,” the study authors wrote.
The authors called for future research to study new pretraining techniques focused on medical imaging to improve performance on tasks that “require highly localized information.”
Radiology should welcome foundation models, but with careful evaluation and realistic expectations, wrote Shandong Wu, PhD, from the University of Pittsburgh, in an accompanying editorial. Wu suggested that with refinement, these models could become valuable tools for breast imaging and for radiology in general.
“DBT-DINO moves the field closer to that goal by showing both the promise and the remaining work required to translate foundation models from computational innovation into clinically meaningful radiology tools,” he wrote.
Read the full study here.