Resting-State Functional MRI Adaptation with Attention Graph Convolution Network for Brain Disorder Identification.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Multi-site resting-state functional magnetic resonance imaging (rs-fMRI) data aids machine learning model training but suffers from significant site-specific data heterogeneity, hindering generalization.
• Graph Convolutional Networks (GCNs) show promise for identifying fMRI biomarkers but often overlook the varying importance of different brain regions (ROIs) in disease diagnosis.

Purpose of the Study:

• To develop a novel multi-site rs-fMRI adaptation framework, termed Attention GCN (A²GCN), for improved brain disorder identification.
• To address data heterogeneity across imaging sites and enhance the interpretability of GCNs by incorporating ROI contributions.

Main Methods:

• The A²GCN framework integrates three key components: GCN-based node representation learning for rs-fMRI features, a node attention mechanism to weigh ROI importance, and a domain adaptation module using mean absolute error and covariance constraints.
• Domain adaptation techniques were employed to mitigate distribution differences between imaging sites.

Main Results:

• The proposed A²GCN framework successfully reduced data heterogeneity across multiple imaging sites.
• The method demonstrated improved performance in the automated recognition of autism spectrum disorders using fMRI data from the ABIDE database.
• The attention mechanism enhanced the interpretability by identifying significant ROIs.

Conclusions:

• The A²GCN framework offers a robust solution for analyzing multi-site rs-fMRI data, effectively handling data heterogeneity.
• This approach improves the accuracy and interpretability of machine learning models for brain disorder identification, particularly for autism spectrum disorder.