Federated learning for multi-center imaging diagnostics: a simulation study in cardiovascular disease

Area of Science:

• Medical imaging
• Artificial intelligence
• Cardiology

Background:

• Deep learning for disease diagnosis is limited by medical data scarcity and heterogeneity.
• Single-center studies lack generalizability, hindering reliable automated diagnosis.
• Federated learning offers a privacy-preserving solution for multi-center medical studies.

Purpose of the Study:

• To evaluate federated learning for hypertrophic cardiomyopathy diagnosis using cardiovascular magnetic resonance (CMR).
• To assess the impact of shape priors and data augmentation on federated learning performance.
• To compare federated learning with traditional centralized learning in a multi-center setting.

Main Methods:

• Simulated a federated learning study using four centers from M&M and ACDC CMR datasets.
• Adapted a 3D-CNN model, incorporating shape priors and various data augmentation techniques.
• Systematically analyzed the effects of different collaborative learning strategies.

Main Results:

• Federated learning achieved promising results competitive with centralized learning, despite small dataset size (180 subjects).
• Federatively trained models demonstrated increased robustness compared to centralized approaches.
• Federated models showed higher sensitivity to domain shift effects.

Conclusions:

• Federated learning is a viable and effective privacy-preserving method for multi-center CMR-based disease diagnosis.
• The approach enhances model robustness and generalizability across different institutions.
• Further research can optimize federated learning strategies for complex medical imaging tasks.