LOCALIZING ACTIVITY SOURCE IN THE BRAIN USING MULTI-DIMENSIONAL WAVELET NEURAL NETWORK

A multi-dimensional wavelet neural network is proposed to localize current dipole within the brain based on a head volume conductor model of realistic shape. Training patterns containing 127 scalp potential elements in response to current dipoles are calculated by using the boundary element method. We utilize scalp potentials and dipole parameters, respectively, as the inputs and outputs to the net. A total of 8,000 pairs of input and output vectors are generated by the boundary element method to train the net. After training, 2,000 independent pairs of test vectors are utilized to evaluate dipole localization results. Our data indicate that the wavelet neural network is able to perform dipole localization both efficiently and accurately.