Enhancing Brain Source Reconstruction by Initializing 3-D Neural Networks With Physical Inverse Solutions.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Accurate brain source reconstruction from electroencephalography (EEG) is vital for understanding brain function and dysfunction.
• Traditional EEG source localization methods often rely on manual priors, lacking data-driven flexibility.
• Current deep learning approaches typically use forward models solely for training data generation, not for direct integration.

Purpose of the Study:

• To develop a novel hybrid method, 3D-PIUNet, for improved EEG source localization.
• To integrate the strengths of traditional physics-based methods and data-driven deep learning.
• To enhance the spatial accuracy of identifying brain activity origins from EEG signals.

Main Methods:

• A hybrid approach, 3D-PIUNet, was developed, starting with a physics-informed estimate using the pseudo-inverse.
• A 3D convolutional U-Net was employed to process the brain as a 3D volume, capturing spatial dependencies.
• The model was trained on simulated pseudo-realistic brain source data with diverse source distributions.

Main Results:

• 3D-PIUNet demonstrated significantly improved spatial accuracy in EEG source localization compared to traditional and end-to-end deep learning methods.
• The method successfully identified the visual cortex and reconstructed temporal activity in real EEG data from a visual task.
• Validation with real-world data confirmed the practical applicability and superior performance of the proposed technique.

Conclusions:

• The 3D-PIUNet method offers a powerful and accurate solution for EEG source localization.
• This hybrid approach effectively bridges the gap between physics-based and data-driven techniques in neuroscience.
• The findings highlight the potential of 3D-PIUNet for advancing brain imaging and understanding neurological processes.