Fast robust subject-independent magnetoencephalographic source localization using an artificial neural network

Area of Science:

• Neuroscience
• Biophysics
• Computational Neuroscience

Background:

• Magnetoencephalography (MEG) is a non-invasive neuroimaging technique.
• Accurate real-time source localization is crucial for understanding brain activity.
• Previous methods required subject-specific retraining, limiting efficiency.

Purpose of the Study:

• To develop a real-time system for accurate single dipole localization using MEG data.
• To overcome the need for subject-specific retraining by incorporating head position.
• To improve the speed and accuracy of dipole localization compared to existing methods.

Main Methods:

• A multilayer perceptron (MLP) was trained to map MEG sensor signals and head position to dipole location.
• The training dataset was generated using an analytic model with simulated dipole locations and real MEG noise.
• A Levenberg-Marquardt optimization routine was used with MLP output as an initial guess.

Main Results:

• The MLP system achieved real-time localization in 0.7 ms with an average error of 0.90 cm.
• An iterative refinement using Levenberg-Marquardt improved accuracy to 0.53 cm in 15 ms.
• The system demonstrated comparable or superior performance to commercial software for MEG source localization.

Conclusions:

• The developed MLP-based system provides a fast and accurate method for real-time MEG source localization.
• Incorporating head position in the MLP model enhances generalizability across subjects and sessions.
• This approach offers a significant advancement for analyzing neural activity using MEG data.