Localized estimation of electromagnetic sources underlying event-related fields using recurrent neural networks

Area of Science:

• Computational Neuroscience
• Biophysics
• Signal Processing

Background:

• Noninvasive electromagnetic measurements like magnetoencephalography (MEG) offer insights into brain activity.
• Accurate reconstruction of neural activity from these signals is crucial for understanding brain function.
• Existing source reconstruction methods have limitations in biological realism and signal fidelity.

Purpose of the Study:

• To develop and validate a recurrent neural network (RNN) model for reconstructing neural activity from electromagnetic signals.
• To improve the biological realism and accuracy of neural source localization compared to traditional methods.
• To leverage biophysical constraints within the RNN framework for enhanced signal reconstruction.

Main Methods:

• Utilized an RNN with fixed output weights representing the lead field matrix derived from boundary element method and co-registered MRI/MEG data.
• Trained the RNN to minimize mean-squared error between outputs and MEG signals, followed by L1 regularization for focused activations.
• Validated the approach using simulations and real MEG data from an auditory oddball experiment, comparing with beamformers, minimum-norm estimate, and mixed-norm estimate.

Main Results:

• The proposed RNN method demonstrated higher output signal-to-noise ratios.
• Achieved comparable correlation and error metrics between estimated and simulated neural sources.
• Reconstructed MEG signals showed superior or equal similarity to ground-truth compared to other methods.
• Source signals estimated from real MEG data were biophysically plausible and consistent with existing literature.

Conclusions:

• The RNN approach, incorporating biophysical constraints, represents a significant advancement in neural source reconstruction.
• This method offers greater biological realism for analyzing event-related neural responses.
• It opens new possibilities for investigating the impact of input manipulations on localized neural activity.