Benchmarking M/EEG inverse solutions to noise level misspecifications

1644 

Abstract Submission 

Anuja Negi^1,2, Stefan Haufe^1,2,3,4, Alexandre Gramfort⁵, Ali Hashemi^1,6

¹Technical University Berlin, Berlin, Germany, ²Bernstein Center for Computational Neuroscience Berlin, Berlin, Germany, ³Charité – Universitätsmedizin Berlin, Berlin, Germany, ⁴Physikalisch - Technische Bundesanstalt Institute Berlin, Berlin, Germany, ⁵Inria, Université Paris-Saclay, Paris, Paris, ⁶The Berlin Institute for the Foundations of Learning and Data (BIFOLD), Berlin, Germany

Anuja Negi  
Technical University Berlin|Bernstein Center for Computational Neuroscience Berlin
Berlin, Germany|Berlin, Germany

Stefan Haufe  
Technical University Berlin|Bernstein Center for Computational Neuroscience Berlin|Charité – Universitätsmedizin Berlin|Physikalisch - Technische Bundesanstalt Institute Berlin
Berlin, Germany|Berlin, Germany|Berlin, Germany|Berlin, Germany

Alexandre Gramfort  
Inria, Université Paris-Saclay
Paris, Paris

Ali Hashemi  
Technical University Berlin|The Berlin Institute for the Foundations of Learning and Data (BIFOLD)
Berlin, Germany|Berlin, Germany

Magneto- and electroencephalographic (M/EEG) are non-invasive techniques that can measure electrical activity in the brain. Brain source imaging (BSI) is used to infer the underlying brain activity from the M/EEG signals. However, this is a challenging ill-posed inverse problem. BSI methods are susceptible to inaccurate source localization, some more than others. This can result from small degrees of under- or overfitting that heavily depend on the choice of modeling parameters.

To compare the feasibility of inverse estimators at varying regularization strengths or estimated noise levels, we benchmark their performance to study the ranges of noise misspecification within which different BSI approaches can still localize well.

We conduct extensive simulations in realistic MEG volume conductor models based on the CamCan dataset [5]. One to five dipolar sources with orientation normal to the cortex are randomly positioned and are assigned a Gaussian random time course of length 10. Sources are then mapped to MEG channels, where Gaussian white noise is added to yield varying signal-to-noise ratios (SNR) ranging from -20 to 40 dB. Sources are reconstructed using smooth linear inverse solutions and sparse non-linear learning solutions [1-4, 6-8] for predefined choices of noise variance for each SNR.

For each reconstruction, performance is evaluated using earth-mover's distance (Fig. 1) and other metrics like euclidean distance, mean squared error, etc. Further, we also assess how noise learning and cross-validation (CV) approaches approximate the "sweet spot" leading to optimal localization. It is observed that most methods work best at high SNR levels and have varying SNR points at which their performance begins to drop. For a fixed SNR, best performance is usually observed with a moderate degree of underfitting. Furthermore, noise variance picked by spatial CV fitting results in a reconstructed SNR close to the ground truth SNR and near-optimal localization performance.

This work allows for a unified and principled framework to compare the performance of BSI methods. Our results contribute to the understanding of BSI techniques by providing a comprehensive evaluation of different inverse estimators and their performance under varying conditions and highlights the importance of carefully selecting modeling parameters in BSI methods. All methods and experiments are publicly available within the "BSI zoo" (github.com/braindatalab/BSI-Zoo) python package, to facilitate further investigation of novel BSI techniques.

Bayesian Modeling

EEG/MEG Modeling and Analysis ¹

Methods Development ²

Other Methods

Computational Neuroscience

Electroencephaolography (EEG)

MEG

Modeling

Open-Source Software

Source Localization