Towards generalizable Federated Learning in medical imaging: A real-world case study on mammography data

Area of Science:

• Medical AI
• Machine Learning
• Computer Science

Background:

• Federated learning (FL) is increasingly used in medical applications for enhanced data privacy.
• Direct application of classic AI experiments to FL is challenging.
• Breast cancer classification requires robust and private AI models.

Purpose of the Study:

• To adapt and evaluate federated learning for breast cancer classification.
• To ensure equivalence between classic and federated model training.
• To introduce a robust FL framework for real-world medical data.

Main Methods:

• Comparison of classic AI model training with a federated variant.
• Introduction of a Breast Area Detection tool for data harmonization in FL.
• Development of an end-to-end FL framework for multi-hospital deployment.

Main Results:

• The proposed FL framework demonstrated significant performance improvements at one hospital.
• The framework achieved results comparable to the classic centralized approach.
• Interventions improved model performance by approximately 35%.

Conclusions:

• Federated learning can be effectively adapted for breast cancer classification with appropriate pre-processing.
• The Breast Area Detection tool enhances FL robustness through data harmonization.
• The developed FL framework offers a viable solution for private, high-performance medical AI.