A deep graph neural network architecture for modelling spatio-temporal dynamics in resting-state functional MRI data.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) is crucial for understanding human brain organization.
• Traditional rs-fMRI analysis models the brain as a graph of regions of interest (ROIs) and their associations.
• Graph neural networks (GNNs) show promise for relational data but are underexplored for rs-fMRI's spatio-temporal dynamics.

Purpose of the Study:

• To present a novel deep neural network architecture for end-to-end analysis of rs-fMRI data.
• To integrate spatial and temporal learning by combining GNNs and temporal convolutional networks (TCNs).
• To leverage both intra-feature (temporal dynamics) and inter-feature (ROI interactions) learning.

Main Methods:

• Developed a hybrid deep learning model integrating GNNs and TCNs.
• The model performs intra-feature learning using TCNs for temporal dynamics.
• The model performs inter-feature learning using GNNs for ROI interactions.

Main Results:

• The model was evaluated on large datasets (UK Biobank, HCP) in unimodal and multimodal settings.
• Ablation studies confirmed the model's effectiveness.
• The architecture demonstrated explainability features mapping to neurobiological insights.

Conclusions:

• The proposed GNN-TCN architecture effectively captures the spatio-temporal nature of rs-fMRI data.
• This model offers a powerful tool for analyzing complex brain dynamics.
• It lays the foundation for future deep learning approaches in rs-fMRI research.