VSSI_2p-Net: Physics-guided deep unfolding with L_2p-norm and variation sparsity for EEG source imaging

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Signal Processing

Background:

• Electroencephalogram (EEG) source imaging (ESI) is an underdetermined problem, challenging traditional neuroimaging methods.
• Existing techniques often require manual parameter tuning for optimal prior information integration.
• Deep learning methods offer data-driven parameter optimization but lack interpretability and require large datasets.

Purpose of the Study:

• To propose a novel neural network model, VSSI2p-Net, integrating advantages of traditional and deep learning ESI methods.
• To address the challenge of parameter optimization and interpretability in ESI.
• To achieve more accurate and efficient ESI solutions.

Main Methods:

• Developed a deep unfolding neural network model named VSSI2p-Net.
• Introduced variation sparsity and ℓ2,p norm (0
• Utilized the Alternating Direction Method of Multipliers (ADMM) for iterative solving and mapped it to a neural network for end-to-end parameter optimization.

Main Results:

• VSSI2p-Net demonstrated superior performance compared to traditional and state-of-the-art deep learning methods on synthetic and real datasets.
• Significant improvements were observed in source localization accuracy and spatial range estimation.
• The proposed method also showed enhanced imaging speed across various source configurations.

Conclusions:

• VSSI2p-Net offers a flexible integration of prior information while maintaining interpretability, outperforming existing ESI methods.
• The model provides a more accurate and efficient solution for the underdetermined ESI problem.
• This approach advances the field of EEG source imaging by addressing key limitations of current techniques.