How Forgiving are M/EEG Inverse Solutions to Noise Level Misspecification? An Excursion into the BSI-Zoo

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biophysics

Background:

• Brain source imaging (BSI), or source localization, is crucial for interpreting magneto- and electroencephalographic (M/EEG) data.
• It is an ill-posed inverse problem sensitive to modeling and experimental parameters like regularization and noise.
• Suboptimal parameter choices can lead to inaccurate source estimation (under- or overfitting).

Purpose of the Study:

• To investigate the robustness of different BSI methods to noise misspecification.
• To compare the performance of linear versus non-linear Bayesian BSI approaches.
• To evaluate noise estimation and cross-validation techniques for hyperparameter selection.

Main Methods:

• Extensive simulations of brain sources with varying sensor noise levels.
• Comparison of smooth linear inverse solutions and sparse non-linear Bayesian learning solutions using Earth Mover's Distance (EMD).
• Assessment of noise estimation and spatial cross-validation for hyperparameter tuning.

Main Results:

• Characterized the robustness of common BSI methods to noise and regularization misspecification.
• Found that moderate underfitting to noise generally yielded better localization performance than overfitting.
• Demonstrated the effectiveness of spatial cross-validation in identifying near-optimal hyperparameters.
• Made BSI methods and experiments available via the BSI-Zoo Python package.

Conclusions:

• The selection of BSI method and hyperparameter estimation significantly impacts localization accuracy and robustness.
• Provides practical guidance for optimizing BSI method selection and tuning.
• Aims to enhance the reliability and reproducibility of M/EEG source localization.