From pretraining to privacy: federated ultrasound foundation model with self-supervised learning.

This study introduces UltraFedFM, a privacy-preserving ultrasound foundation model developed through federated learning across 16 medical institutions in 9 countries.

The model was trained using over 1 million ultrasound images covering 19 organs and 10 ultrasound modalities.

Key advantages of UltraFedFM include:

• Reduced reliance on large centralized labeled datasets.
• Expanded task versatility beyond existing AI ultrasound methods.
• Strong generalization, achieving an average AUROC of 0.927 for disease diagnosis.
• High lesion segmentation accuracy with a dice similarity coefficient of 0.878.
• Performance surpasses mid-level sonographers and matches expert-level accuracy in diagnosing eight common systemic diseases.
• Maintains patient privacy while improving clinical ultrasound diagnostics.