Site Common Information-Guided Site-to-Individual-to-Global Multi-View GCN for Psychiatric Diagnosis

Area of Science:

• Neuroimaging and computational neuroscience
• Psychiatric disorder diagnosis and assessment
• Machine learning applications in healthcare

Background:

• Magnetic resonance neuroimaging (MRI)-based functional connections show promise for psychiatric disorder diagnosis.
• Multi-center neuroimaging studies are crucial for robust findings but face challenges with inter-site variability.
• Existing methods struggle with data consistency and generality due to scanner and protocol differences across sites.

Purpose of the Study:

• To propose a novel framework, sigGCN, to address inter-site variability in multi-center neuroimaging data.
• To harmonize multi-center data for improved diagnostic accuracy in psychiatric disorders.
• To enhance the reliability and generality of neuroimaging-based diagnostic models.

Main Methods:

• Developed a site common information guided Site-to-Individual-to-Global Multi-View graph convolutional network (sigGCN) framework.
• SigGCN extracts shared site information to mitigate site-effects and constructs individual multi-view representations.
• Integrated individual and site-common features to generate global Chebyshev networks for analysis.

Main Results:

• SigGCN achieved the highest accuracy in diagnosing autism compared to existing models.
• Successfully classified autism and schizophrenia from control groups across multiple datasets and brain atlases.
• Demonstrated robustness and stability in harmonizing multi-center neuroimaging data.

Conclusions:

• The sigGCN framework effectively harmonizes multi-center neuroimaging data, mitigating site-effects.
• SigGCN significantly improves the performance of neuroimaging-based psychiatric disorder diagnosis.
• Offers a reliable approach for leveraging hierarchical information across sites, individuals, and global networks.