Electromagnetic Source Imaging With a Combination of Sparse Bayesian Learning and Deep Neural Network.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Accurate brain activity reconstruction from electroencephalography (EEG) and magnetoencephalography (MEG) is challenging due to the ill-posed nature of the inverse problem.
• Existing methods often struggle with accuracy and robustness in complex scenarios.

Purpose of the Study:

• To develop a novel data-driven source imaging framework to address the challenges in E/MEG inverse problems.
• To enhance the accuracy and robustness of brain activity reconstruction.

Main Methods:

• Proposed a sparse Bayesian learning and deep neural network (SI-SBLNN) framework.
• Compressed variational inference using a deep neural network mapping from measurements to latent sparseness encoding parameters.
• Trained the network with synthesized data from a probabilistic graphical model, utilizing source imaging based on spatio-temporal basis function (SI-STBF) as a backbone.

Main Results:

• Validated the framework's availability across different head models and robustness against varying noise intensities in numerical simulations.
• Demonstrated superior performance compared to SI-STBF and other benchmarks in diverse source configurations.
• Achieved concordant results with prior studies in real-data experiments.

Conclusions:

• The proposed SI-SBLNN framework offers a significant advancement in E/MEG source imaging.
• The data-driven approach enhances accuracy, robustness, and applicability for reconstructing brain activity.