Probabilistic Structure Learning for EEG/MEG Source Imaging With Hierarchical Graph Priors.

Area of Science:

• Neuroscience
• Biophysics
• Signal Processing

Background:

• Brain source imaging (ESI) noninvasively characterizes brain activity using EEG/MEG.
• Traditional ESI methods often ignore temporal structures, increasing noise sensitivity and limiting flexibility.
• Existing methods struggle to account for variations in brain activation over time.

Purpose of the Study:

• To develop a novel probabilistic ESI model addressing noise and temporal variations.
• To enhance spatiotemporal continuity and flexibility in brain activation patterns.
• To improve the accuracy of source localization in EEG/MEG data.

Main Methods:

• Proposed a probabilistic ESI model with a hierarchical graph prior.
• Incorporated a spanning tree constraint for spatiotemporal continuity.
• Developed an efficient alternating convex search algorithm for model optimization.

Main Results:

• Demonstrated significant improvements in source localization performance, particularly at high signal-to-noise ratios (SNRs).
• Validated the method using synthetic data on a realistic brain model and real EEG/MEG datasets.
• Achieved neurologically plausible ESI reconstructions in real-world applications.

Conclusions:

• The novel probabilistic ESI model effectively handles noise and captures temporal dynamics.
• The proposed method offers superior source localization compared to benchmark techniques.
• This approach advances noninvasive brain activity characterization using EEG/MEG.