A network inversion technique for estimating equivalent dipole description of visual evoked potential

For the activation study of the brain, dipole localization from the scalp potential is one of the most promising techniques to realize a reasonable temporal resolution which cannot be realized in functional MR and PET. The goal of our study is to estimate inversely the electrical brain activity in the form of several dipoles from the scalp potential, using a network inversion technique. As a basic approach, we have inversely estimated several dipoles from the potential distribution on a spherical surface, in the homogeneous sphere model. In the training phase, by expanding the neural network input dimensions being redundant, the network can easily learn the forward mapping. In the inversion phase, the space of the expanded-network-input-vector can be narrowed by introducing a penalty term. Additionally, a consensus term was used to force several dipoles to have a similar orientation. We estimate that this is applicable to the localization of several dipoles that reflect the actual brain activity, especially in the visual evoked potentials.