ADMM-ESINet: A Deep Unrolling Network for EEG Extended Source Imaging.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computational Neuroscience

Background:

• Electroencephalography (EEG) source imaging (ESI) is vital for brain research and diagnosing disorders.
• Traditional ESI methods face real-time reconstruction challenges.
• Deep neural network (DNN) ESI methods often lack generalization.

Purpose of the Study:

• To develop a robust and efficient deep unfolding neural network for real-time EEG source imaging.
• To improve the generalization ability of DNN-based ESI methods.
• To accurately reconstruct the location, extent, and temporal dynamics of extended cortical sources.

Main Methods:

• Proposed ADMM-ESINet, a deep unfolding neural network integrating structured sparsity and the Alternating Direction Method of Multipliers (ADMM).
• Unrolled the ADMM algorithm into a cascaded network architecture for end-to-end learning.
• Learned regularization parameters and spatial transform operators directly from training data.

Main Results:

• ADMM-ESINet demonstrated superior generalization compared to traditional DNN-based methods.
• Achieved accurate reconstruction of extended EEG sources, including location, extent, and temporal dynamics.
• Enabled real-time ESI reconstruction.

Conclusions:

• ADMM-ESINet is a promising method for robust and efficient real-time EEG source imaging.
• The approach effectively integrates prior knowledge into a deep learning framework.
• The method advances the capabilities for understanding brain function and disorders using EEG data.