EEG Source Imaging using GANs with Deep Image Prior.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Medical Imaging

Background:

• Noninvasive brain source localization using electroencephalogram (EEG) is challenging.
• Traditional methods rely on handcrafted regularization terms and neural-physiological assumptions.
• Deep learning shows promise for solving inverse problems in medical imaging.

Purpose of the Study:

• To propose a novel unsupervised learning, training-free framework for EEG source localization.
• To leverage Generative Adversarial Networks and deep image prior (GANs-DIP) for simulating spatially structured source signals.
• To faithfully recover extended source activation patterns in the brain.

Main Methods:

• Developed a GANs-DIP framework as a generative model.
• Simulated spatially structured source signals.
• Applied the framework in an unsupervised manner for source recovery.

Main Results:

• The proposed framework successfully recovered extended source patches activation patterns.
• Experiments were conducted on a realistic brain model with varying signal-to-noise ratios (SNRs).
• Satisfactory performance was observed in recovering underlying source activation.

Conclusions:

• The GANs-DIP framework offers a promising approach for unsupervised EEG source localization.
• This training-free method accurately reconstructs brain activity patterns.
• The approach demonstrates robustness across different SNR levels.