Knowledge-Aware Multisite Adaptive Graph Transformer for Brain Disorder Diagnosis.

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Diagnosing brain disorders using resting-state functional magnetic resonance imaging (rs-fMRI) is challenging due to complex imaging features and limited sample sizes.
• Graph convolutional networks (GCNs) show promise in brain disorder diagnosis by modeling individual and population interactions but face limitations.
• Existing GCN methods overlook feature sensitivity to non-imaging data, inter-feature relationships, and struggle with multisite data heterogeneity.

Purpose of the Study:

• To propose a knowledge-aware multisite adaptive graph Transformer to overcome limitations in GCN-based brain disorder diagnosis.
• To improve the accuracy and robustness of brain disorder classification using rs-fMRI data from multiple sites.

Main Methods:

• Constructed feature-sensitive and feature-insensitive subgraphs by evaluating feature sensitivity to non-imaging information.
• Integrated a Transformer module to capture intrinsic relationships between features after subgraph fusion.
• Employed a domain-adaptive GCN with multiple loss functions to mitigate inter-site data heterogeneity for classification.

Main Results:

• The proposed framework demonstrated state-of-the-art performance on two brain disorder diagnostic tasks.
• Effectively addressed limitations of previous GCN approaches by considering feature sensitivity and inter-feature relationships.
• Successfully mitigated the impact of inter-site heterogeneity in multisite rs-fMRI datasets.

Conclusions:

• The knowledge-aware multisite adaptive graph Transformer offers a robust and accurate approach for brain disorder diagnosis using rs-fMRI.
• This method advances the application of GCNs in neuroimaging by incorporating feature-level and site-level adaptive strategies.
• The framework holds significant potential for improving clinical diagnosis and understanding of brain disorders.