Developing a knowledge-guided federated graph attention learning network with a diffusion module to diagnose...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Limited sample sizes in Alzheimer's disease (AD) research hinder intelligent diagnostic system performance.
• Multi-site data enhances sample size but introduces privacy and heterogeneity challenges.

Purpose of the Study:

• To develop a privacy-preserving, knowledge-guided federated learning network for improved AD diagnosis from multi-site structural MRI data.
• To address data heterogeneity and privacy concerns in collaborative neuroimaging studies.

Main Methods:

• Extracted ROI-based volume features from sMRI data using multiple templates.
• Integrated prior AD knowledge to identify discriminative imaging features.
• Employed an attention-guided diffusion module for data augmentation within a federated learning framework.
• Utilized a federated graph attention learning network for AD classification.

Main Results:

• Validated the approach on three AD datasets, demonstrating improved diagnostic accuracy.
• The federated learning framework successfully maintained inter-site data privacy.
• The diffusion module effectively synthesized samples by prioritizing key AD features, mitigating heterogeneity.

Conclusions:

• The developed knowledge-guided federated graph attention network is a promising tool for optimizing multi-site neuroimaging data.
• This approach significantly enhances the accuracy of Alzheimer's disease diagnosis in clinical settings.
• The study highlights the potential of federated learning for secure and effective collaborative medical research.