FedFound: a federated foundation model for lifespan brain morphological connectome analysis

Area of Science:

• Neuroimaging
• Machine Learning
• Radiology

Background:

• Brain morphological connectomes from structural MRI reveal individual variability and abnormalities.
• Subtle and complex alterations pose challenges for accurate machine learning-based diagnosis.

Purpose of the Study:

• To introduce FedFound, the first federated foundation model for analyzing lifespan brain morphological connectomes.
• To enable robust, scalable, and generalizable diagnosis across diverse neurological disorders.

Main Methods:

• Integration of heterogeneous neuroimaging datasets (22,911 subjects, ages 0-100) from multiple sites and disorders.
• Combination of self-supervised pre-training and supervised federated disease-specific refinement.
• Distributed optimization for multidisciplinary knowledge aggregation.

Main Results:

• FedFound achieved superior performance and interpretability across nine diagnostic tasks (neurodevelopmental, neuropsychiatric, neurodegenerative disorders).
• The model revealed shared and disorder-specific morphological patterns across different etiologies.
• Demonstrated robust and scalable analysis of lifespan brain connectomes.

Conclusions:

• FedFound offers a robust foundation for lifespan neuroimage-based diagnosis, complementing clinical expertise.
• Establishes a scalable paradigm for integrating heterogeneous neuroimaging data across institutions, populations, and diseases.
• Advances the development of medical foundation models for brain imaging analysis.