Deep Residual Neural Networks for Spatial EEG Source Imaging.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computational Neuroscience

Background:

• Non-invasive study of brain function relies heavily on electroencephalography (EEG) source imaging.
• Traditional EEG inverse problem solutions face challenges due to the ill-posed nature of mapping scalp potentials to cortical sources.
• Similar scalp EEG patterns can arise from different underlying brain activation, complicating accurate source localization.

Purpose of the Study:

• To introduce DeepMapper, a novel framework for enhanced EEG spatial source imaging.
• To address the ill-posed nature of the EEG inverse problem using a deep learning approach.
• To improve the accuracy and reliability of non-invasive brain activity localization.

Main Methods:

• Developed a two-part framework: dataset simulation and neural network training.
• Utilized a 3-shell realistic head model and boundary element method (BEM) for EEG forward problem simulation.
• Incorporated cortical functional atlases for physiological constraints and simulated extensive EEG-cortex data for supervised learning.

Main Results:

• DeepMapper demonstrated superior performance in recovering EEG spatial sources compared to traditional methods.
• The neural network effectively stored prior information through supervised learning in weight layers and nonlinear connections.
• Simulation results validated the efficacy of the proposed deep learning framework.

Conclusions:

• DeepMapper offers a novel and effective approach to EEG spatial source imaging.
• The method shows significant potential for advancing non-invasive brain function studies.
• DeepMapper provides a more accurate solution for the challenging EEG inverse problem.