Dynamic Graph Representation Learning for Spatio-Temporal Neuroimaging Analysis.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Neuroimaging analysis seeks to understand brain function noninvasively.
• Graph neural networks (GNNs) capture brain network structures but often overlook temporal dynamics.
• Existing methods primarily focus on spatial functional connectivity, neglecting complex temporal patterns.

Purpose of the Study:

• To propose a spatio-temporal interactive graph representation framework (STIGR) for dynamic neuroimaging analysis.
• To enhance the capture of interrelationships between spatial and temporal brain dynamics.
• To provide a versatile and interpretable tool for medical professionals in neuroimaging research.

Main Methods:

• Leveraging a dynamic adaptive-neighbor graph convolution network to model spatio-temporal dynamics.
• Incorporating a self-attention module based on Transformers to capture long-term dependencies.
• Utilizing contrastive learning to model cross-temporal correlations in dynamic graphs.

Main Results:

• Achieved state-of-the-art performance in classification and regression tasks across six public neuroimaging datasets.
• Demonstrated competitive performance of STIGR across different platforms.
• Enabled detection of significant temporal association patterns in neuroimaging signals.

Conclusions:

• STIGR offers a powerful approach for dynamic neuroimaging analysis, integrating spatial and temporal information effectively.
• The framework provides a versatile and interpretable tool for identifying neurological patterns.
• This work advances the field by addressing the limitations of existing methods in capturing temporal brain dynamics.