Deep source transfer learning for the estimation of internal brain dynamics using scalp EEG

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Electroencephalography (EEG) offers high temporal resolution but limited spatial resolution for brain activity.
• Source imaging techniques aim to improve EEG spatial resolution for better neural decoding and brain-computer interaction (BCI).

Purpose of the Study:

• To develop a novel, data-driven EEG source imaging scheme using deep learning for precise estimation of macroscale spatiotemporal brain dynamics.
• To enhance the reliability and accuracy of neural decoding and BCI applications.

Main Methods:

• A deep source imaging framework utilizing a convolutional-recurrent neural network was designed for high-density EEG recordings.
• A comprehensive brain model with 210 cortical regions and 16 thalamic nuclei was created for synthetic data generation.
• Transfer learning was applied to bridge the gap between synthetic and realistic EEG data.

Main Results:

• The proposed deep learning method accurately estimated spatial and temporal brain source activity.
• The framework demonstrated superior performance compared to existing state-of-the-art approaches.
• The method successfully located seizure onset zones in epilepsy and reconstructed thalamocortical interactions during acupuncture stimulation.

Conclusions:

• The developed EEG data-driven source imaging framework offers precise and efficient estimation of brain dynamics.
• This approach holds significant potential for advancing neuroscience research and clinical applications, including BCI.
• The method's effectiveness in epilepsy and sensory processing highlights its broad applicability.