Related Experiment Videos
Fast accurate MEG source localization using a multilayer perceptron trained with real brain noise
Sung Chan Jun1, Barak A Pearlmutter, Guido Nolte
1Department of Computer Science, University of New Mexico, Albuquerque 87131, USA. junsc@cs.unm.edu
Physics in Medicine and Biology
|August 13, 2002
Summary
A new hybrid method uses a trained multilayer perceptron (MLP) to initialize the Levenberg-Marquardt (LM) algorithm, significantly speeding up source localization for electroencephalographic (EEG) and magnetoencephalographic (MEG) signals.
Area of Science:
- Neuroscience
- Biophysics
- Computational Science
Background:
- Iterative gradient methods like Levenberg-Marquardt (LM) are standard for source localization using EEG/MEG.
- LM's sensitivity to initial guesses and high computational cost necessitate improvements.
Purpose of the Study:
- To develop a faster and more accurate source localization method for EEG/MEG signals.
- To reduce the computational burden of traditional LM algorithms.
Main Methods:
- Trained a multilayer perceptron (MLP) to perform real-time source localization.
- Used an analytical model of electromagnetic propagation for training data generation.
- Developed a hybrid MLP-start-LM method, initializing LM with MLP outputs.
- Compared localization performance of LM, MLPs, and hybrid methods with varying noise conditions.
Main Results:
- Trained MLPs offer significant speed improvements over traditional methods.
- Hybrid MLP-start-LM methods are faster and more accurate than standard LM with multiple initial guesses.
- MLP trained with real brain noise provides an initialization comparable to the true dipole location for LM.
Conclusions:
- Hybrid MLP-start-LM significantly enhances the speed and accuracy of EEG/MEG source localization.
- MLP-based initialization effectively addresses LM's sensitivity to initial guesses.
- The developed hybrid method shows near-optimal performance, making real-time source localization feasible.