dFCExpert: Learning Dynamic Functional Connectivity Patterns with Modularity and State Experts.

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Characterizing dynamic functional connectivity (dFC) from fMRI is crucial for understanding brain function and disorders.
• Existing graph neural network (GNN) models struggle with brain modularity and varying dFC states.

Purpose of the Study:

• To introduce dFCExpert, a novel method for robust dFC pattern representation learning.
• To address limitations in modeling brain modularity and dFC state variations.

Main Methods:

• Combines GNNs with Mixture of Experts (MoE) for modularity experts, focusing on functional network modules.
• Employs state experts with soft prototype clustering to identify distinct dFC states.
• Utilizes two large-scale fMRI datasets for validation.

Main Results:

• dFCExpert demonstrates superior performance compared to existing methods.
• The learned dFC representations offer improved interpretability.
• The method shows promise for enhancing clinical diagnosis of brain disorders.

Conclusions:

• dFCExpert effectively models brain modularity and dynamic connectivity states.
• The approach provides interpretable and clinically relevant insights into brain function.
• This method advances the analysis of fMRI data for neuroscience and medicine.