Source reconstruction accuracy of MEG and EEG Bayesian inversion approaches

Area of Science:

• Neuroscience
• Biophysics
• Computational Biology

Background:

• Electroencephalography (EEG) and magnetoencephalography (MEG) offer high temporal resolution for studying brain activity.
• Accurate source localization is critical for network detection using EEG/MEG data.
• Existing source localization methods vary in their underlying assumptions and performance.

Purpose of the Study:

• To quantitatively compare four source localization schemes within a Variational Bayesian framework.
• To evaluate the performance of Minimum Norm Model (MNM), Empirical Bayesian Beamformer (EBB), Automatic Relevance Determination (ARD), and Greedy Search (GS).
• To assess the impact of source number, signal-to-noise ratio (SNR), and temporal source correlation on localization accuracy.

Main Methods:

• Implementation of MNM, EBB, ARD, and GS within a unified Variational Bayesian framework.
• Quantitative comparison of localization performance across varying numbers of sources (1-3).
• Evaluation under different signal-to-noise ratios (SNRs) and temporal correlations of source activity.

Main Results:

• MNM is effective only for single-source configurations.
• EBB achieves millimeter accuracy with high SNRs and low source correlation.
• ARD and GS demonstrate greater robustness to noise and temporal correlations, with centimeter-level accuracy.

Conclusions:

• ARD and GS provide more reliable network detection in complex EEG/MEG scenarios compared to MNM and EBB.
• The choice of source localization method significantly impacts the accuracy of brain network analysis.
• Further investigation into anatomical priors did not improve ARD/GS performance in this study.