DBT-DINO: Domain-Specific Pretraining Enhances Breast Density Classification but Not Cancer Risk or Lesion Detection

A DBT foundation model (DBT-DINO) improved breast density classification accuracy (79% vs 73%, p<.001) but did not significantly outperform ImageNet pretraining for 5-year breast cancer risk prediction (AUC 0.78 vs 0.76) or lesion detection (sensitivity 62% vs 67%).

• For 5-year breast cancer risk prediction, DBT-DINO AUC was 0.78 vs 0.76 for baseline (P=.057), no evidence of difference.
• Lesion detection sensitivity: DBT-DINO 62% (84/136 lesions) vs baseline 67% (91/136), P=.60.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.