EEG Source Analysis with a Convolutional Neural Network and Finite Element Analysis.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computational Neuroscience

Background:

• Source analysis of electrophysiological brain activity involves solving ill-posed inverse problems.
• Existing algorithms often rely on suboptimal forward modeling for training neural networks.
• There is a need for improved methods in electroencephalography (EEG) source analysis.

Purpose of the Study:

• To propose and evaluate a novel Convolutional Neural Network (CNN) for reconstructing EEG brain activity.
• To address limitations in current neural network training for EEG source analysis.
• To demonstrate the efficacy of a CNN trained with a realistic head model.

Main Methods:

• Developed a CNN architecture for EEG source analysis.
• Utilized a skull-conductivity calibrated, white matter anisotropic head model for generating simulated EEG data.
• Trained the CNN using the generated simulated EEG data.
• Evaluated CNN performance on simulated data and a real-world somatosensory evoked potential experiment.

Main Results:

• The CNN successfully reconstructed EEG brain activity.
• Accurate localization of the P20/N20 component at Brodmann area 3b was achieved.
• The CNN demonstrated potential for localizing deeper brain sources.
• Localization performance was comparable to established methods like single dipole scans and sLORETA.

Conclusions:

• The proposed CNN offers a promising approach for EEG source analysis.
• Realistic head modeling is crucial for effective neural network training in this domain.
• The CNN shows potential for real-world clinical and research applications in brain activity reconstruction.