Efficient federated learning for distributed neuroimaging data.

Area of Science:

• Neuroscience
• Computer Science
• Data Science

Background:

• Neuroimaging research increasingly involves data sharing.
• Institutional data control due to privacy and accountability concerns hinders collaboration.
• A need exists for tools enabling analysis of distributed datasets without direct data transfer.

Purpose of the Study:

• To propose a decentralized sparse federated learning (FL) strategy for analyzing amalgamated neuroimaging datasets.
• To address challenges of data privacy, security, and institutional control in collaborative research.
• To reduce communication overhead in federated learning frameworks.

Main Methods:

• Developed a decentralized sparse federated learning (FL) strategy.
• Emphasized local training of sparse models to minimize data transmission.
• Implemented selective sharing of model parameters between client sites.
• Utilized the Adolescent Brain Cognitive Development (ABCD) dataset for validation.

Main Results:

• The proposed FL strategy significantly lowers communication overheads.
• Efficiency gains are more substantial with larger models and diverse site resources.
• Demonstrated the effectiveness of the approach on a large-scale neuroimaging dataset.

Conclusions:

• Decentralized sparse FL offers an effective solution for collaborative neuroimaging analysis.
• The method enhances efficiency and scalability in federated learning environments.
• This approach facilitates secure and privacy-preserving analysis of sensitive data across institutions.