Functional connectivity learning via Siamese-based SPD matrix representation of brain imaging data.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Brain functional connectivity analysis is crucial for understanding brain dynamics.
• Existing methods often rely on application-specific settings and deterministic models.
• There is a need for automated, application-independent measures of functional connectivity.

Purpose of the Study:

• To propose a novel Siamese-based Symmetric Positive Definite (SPD) Matrix Representation framework (SiameseSPD-MR).
• To develop an application-independent measure for brain imaging data (BID), specifically Electroencephalography (EEG).
• To automatically learn functional connectivity representations from deep features.

Main Methods:

• Utilizing graph convolution to extract features from BID based on anatomical structure.
• Applying an adaptive Gaussian kernel function for functional connectivity representation.
• Employing a Siamese network architecture with SPD matrix transformation for similarity derivation.

Main Results:

• SiameseSPD-MR demonstrated superior ability to capture functional connectivity differences compared to state-of-the-art methods.
• The framework highlighted typical EEG characteristics associated with autism spectrum disorder (ASD).
• Learned functional connectivity representations served as effective markers for brain network analysis and ASD discrimination.

Conclusions:

• SiameseSPD-MR offers a robust and automated approach to functional connectivity analysis in EEG.
• The proposed method provides meaningful markers for understanding brain networks and diagnosing neurological conditions like ASD.
• This framework advances the development of application-independent measures in brain imaging data analysis.