ADHD Classification With GCN via Joint Feature Learning Among Nodes and Edges

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) derived functional connectivity networks (FCNs) are crucial for understanding brain alterations in attention-deficit/hyperactivity disorder (ADHD).
• Existing graph neural network (GNN) methods often overlook edge information and dynamic node feature interdependencies, limiting diagnostic accuracy for ADHD.

Purpose of the Study:

• To develop an advanced graph convolutional network (JNEL-GCN) for enhanced ADHD classification and biomarker discovery by integrating node and edge features.
• To improve the representation of dynamic interdependencies within brain networks for more accurate neuroimaging-based ADHD diagnosis.

Main Methods:

• Constructed dual graph representations: a node graph with amplitude of low-frequency fluctuations (ALFF) and a node-edge relationship matrix, and an edge graph using line graph theory.
• Implemented an alternating feature update mechanism with optimized graph convolutions for hierarchical learning of node-edge relationships across network layers.
• Utilized gradient-based biomarker analysis to identify ADHD-associated brain regions.

Main Results:

• Achieved high classification accuracy for ADHD: 97.3% on the ADHD200 dataset and 97.1% on the ABIDE-I dataset, surpassing current benchmarks.
• Identified significant ADHD-related regions in bilateral limbic and default mode networks, consistent with existing literature.
• Demonstrated the efficacy of the dual-graph approach in capturing dynamic network interactions.

Conclusions:

• The JNEL-GCN framework offers a significant advancement in neuroimaging-based ADHD diagnosis by comprehensively analyzing dynamic brain network interactions.
• The study provides interpretable biomarkers for ADHD, supporting clinical neuroscience applications and future research.
• This dual-graph approach enhances diagnostic performance and deepens the understanding of brain network abnormalities in ADHD.