Understanding Graph Isomorphism Network for rs-fMRI Functional Connectivity Analysis.

Area of Science:

• Neuroscience
• Machine Learning
• Graph Theory

Background:

• Graph neural networks (GNNs) are increasingly applied to functional magnetic resonance imaging (fMRI) data.
• A key limitation of current GNN applications in fMRI is the difficulty in providing neuroscientifically interpretable explanations for classification results.

Purpose of the Study:

• To develop a novel framework for analyzing fMRI data using Graph Isomorphism Networks (GINs).
• To enhance the neuroscientific explainability of GNN-based classification by visualizing important brain regions.

Main Methods:

• Utilized the Graph Isomorphism Network (GIN), a powerful GNN for graph classification.
• Developed a framework that leverages the dual representation of GINs as a graph-space equivalent of convolutional neural networks (CNNs).
• Adapted CNN-based saliency map techniques for GNNs to visualize critical brain regions in fMRI data.

Main Results:

• Successfully applied the GIN framework to classify subject sex using resting-state fMRI (rs-fMRI) data.
• The generated saliency maps demonstrated a high degree of correspondence with established neuroimaging evidence of sex differences in the brain.

Conclusions:

• The proposed GIN framework offers a neuroscientifically interpretable method for analyzing fMRI data.
• This approach advances the application of GNNs in neuroimaging by enabling visualization of brain regions critical for classification tasks.