Multi-Head Graph Convolutional Network for Structural Connectome Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Brain connectivity analysis from diffusion magnetic resonance images (dMRI) is crucial for understanding neurological function.
• Existing machine learning models may not fully capture complex relationships within brain connectivity data.

Purpose of the Study:

• To develop and evaluate a novel machine learning model for brain connectivity classification.
• To assess the model's performance on sex classification using dMRI data.
• To investigate the utility of graph convolutional networks (GCNs) for capturing brain connectome variations.

Main Methods:

• A machine learning model inspired by graph convolutional networks (GCNs) was proposed.
• The model utilizes a parallel GCN mechanism with multiple heads focusing on graph edges and nodes.
• Experiments were conducted on two publicly available datasets: PREVENT-AD (347 subjects) and OASIS3 (771 subjects).

Main Results:

• The proposed GCN-inspired model achieved the highest performance in sex classification.
• It outperformed classical machine learning algorithms and other deep learning methods.
• The model effectively captured complementary and representative features from brain connectivity data.

Conclusions:

• The novel GCN-based model demonstrates superior performance in classifying sex based on brain connectivity.
• This approach offers a promising tool for analyzing brain connectome variations.
• The findings contribute to a better understanding of sex-related differences in the brain for health and disease research.