Deep EEG Superresolution via Correlating Brain Structural and Functional Connectivities.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Electroencephalogram (EEG) offers millisecond-level neural dynamics but suffers from limited spatial resolution, hindering neuroscience research and consumer EEG applications.
• Existing super-resolution (SR) methods struggle with reconstructing high-resolution (HR) EEG due to challenges in modeling electrode relationships and individual subject variability.

Purpose of the Study:

• To propose Deep-EEGSR, a deep learning framework for reconstructing high-resolution (HR) EEG from low-resolution (LR) data.
• To address the limitations of current SR methods by incorporating brain structural and functional connectivities.

Main Methods:

• Developed Deep-EEGSR, a framework combining a compact convolutional network and a filter generation network (FGN).
• Employed graph convolution to adapt to structural connectivity among EEG channels.
• Utilized sample-specific dynamic convolution with FGN-adjusted filters for subject-specific functional connectivity.

Main Results:

• Deep-EEGSR significantly outperformed state-of-the-art methods, reducing normalized mean squared error (NMSE) by 1%-6% and improving signal-to-noise ratio (SNR) by up to 1.2 dB.
• Reconstructed SR EEG demonstrated superior performance in autism spectrum disorder (ASD) discrimination and spatial localization compared to LR EEG.
• The benefits of SR EEG were observed to increase with the scale of super-resolution.

Conclusions:

• Deep-EEGSR offers an effective end-to-end solution for reconstructing HR EEG.
• The framework's ability to integrate connectivity information enhances EEG analysis, particularly for clinical applications like ASD detection.
• This approach holds promise for advancing EEG-based research and neuroengineering applications.