Brain Functional Connectivity Analysis via Graphical Deep Learning.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models, particularly graphical models, show promise for analyzing brain functional connectivity.
• Challenges in brain network analysis include limited sample sizes and intricate inter-regional relationships.

Purpose of the Study:

• To develop a novel graph convolutional network (GCN) framework for brain functional connectivity analysis.
• To improve the classification of cognitive functions based on brain network data.
• To identify brain regions and networks crucial for cognitive performance.

Main Methods:

• A GCN-based framework was developed, integrating region-to-region brain connectivities and subject-subject relationships.
• An affinity subject-subject graph was constructed and analyzed using GCN.
• Laplacian regularization was employed to mitigate overfitting.
• The model was validated using data from the Philadelphia Neurodevelopmental Cohort.

Main Results:

• The proposed GCN framework demonstrated superior performance in classifying individuals with low versus high Wide Range Achievement Test (WRAT) scores compared to competing models.
• Occlusion sensitivity analysis successfully identified cognition-related brain functional networks.
• The findings align with existing research while also revealing novel insights into brain-cognition relationships.

Conclusions:

• Graph convolutional networks (GCNs), when integrated with prior knowledge of brain networks, provide a potent method for detecting significant brain networks and regions linked to cognitive functions.
• This approach enhances our understanding of the neural underpinnings of cognitive abilities.