Combining Spatial Wavelets and Sparse Bayesian Learning for Extended Brain Sources Reconstruction.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Accurate reconstruction of extended cortical activity from M/EEG data is challenging due to the ill-conditioned nature of the problem.
• Existing methods struggle with precise localization, amplitude, and temporal course estimation, particularly for distributed sources.

Purpose of the Study:

• To develop and validate a novel method for reconstructing extended cortical activity using spectral graph wavelets (SGW) on the cortical surface.
• To accurately localize distributed brain sources, estimating their amplitude and time course from M/EEG data.
• To address numerical optimization challenges inherent in source reconstruction.

Main Methods:

• Modeling distributed M/EEG sources using a system of spectral graph wavelets (SGW) defined on the cortical surface.
• Estimating unknown wavelet coefficients via variational or Bayesian formulations, incorporating sparsity-inducing priors like sparse Bayesian learning (SBL).
• Comparing SGW-based approaches with concurrent methods using real M/EEG data and numerical simulations, assessing reconstruction quality with complementary metrics.

Main Results:

• SGW-based methods accurately identify extended cortical sources.
• The combination of SGW with SBL demonstrated superior performance, eliminating the need for hyperparameter tuning and adapting to varying signal-to-noise ratios (SNR).
• This approach yielded robust results across all metrics and showed remarkable performance in reducing depth bias.

Conclusions:

• Spectral graph cortical wavelets are effective for M/EEG source reconstruction, particularly when combined with SBL.
• Accurate localization, depth, amplitude, and time course estimation of brain activity from M/EEG data holds significant potential for clinical applications, such as improving seizure source detection in epilepsy.