BrainGNN: Interpretable Brain Graph Neural Network for fMRI Analysis.

Area of Science:

• Neuroimaging
• Graph Neural Networks
• Biomarker Discovery

Background:

• Identifying brain regions linked to neurological disorders and cognitive tasks is crucial in neuroimaging research.
• Functional magnetic resonance imaging (fMRI) provides valuable data for understanding brain function.
• Existing methods may lack transparency and fail to fully leverage complex brain graph properties.

Purpose of the Study:

• To introduce BrainGNN, a graph neural network (GNN) framework for analyzing fMRI data.
• To discover neurological biomarkers for conditions such as Autism Spectrum Disorder (ASD).
• To enhance transparency in medical image analysis by highlighting salient brain regions.

Main Methods:

• Developed novel ROI-aware graph convolutional (Ra-GConv) layers to integrate topological and functional fMRI information.
• Incorporated ROI-selection pooling layers (R-pool) for transparent identification of important brain regions (nodes).
• Utilized regularization terms (unit loss, topK pooling loss, group-level consistency loss) for flexible ROI selection.

Main Results:

• BrainGNN demonstrated superior performance over alternative fMRI analysis methods across four evaluation metrics on ASD and Human Connectome Project (HCP) datasets.
• The framework successfully identified salient ROIs and community clusters with high correspondence to known ASD biomarkers and HCP task states.
• Hyperparameter choices were investigated, confirming the robustness and effectiveness of the BrainGNN framework.

Conclusions:

• BrainGNN offers a powerful and transparent framework for fMRI analysis and neurological biomarker discovery.
• The method effectively leverages brain graph properties and enhances interpretability in neuroimaging.
• BrainGNN shows significant potential for advancing the understanding and diagnosis of neurological disorders.