Study Of The Dynamic Magnetoencephalography Inverse Problem

There are more and more researches into the dynamic process of the brain neuronal populations in the field of clinical diagnosis and cognitive science. Magnetoencephalography(MEG)funcitonal imaging estimate the amplitude of all possible source locations. When imaging the dynamic process, the relationship between MEG data and source signals in space as well as in time domain should be studied. Based on the widely used minimum norm estimate(MNE), we proposed two methods to solve the dynamic MEG inverse problem.Two-parameter regularization has been adopted in researches on kinds of fields, such as electrocardiograph(ECG) inverse problem, medical image reconstruction,electrical conductivity imaging. We utilized two-parameter regularization to solve the dynamic MEG inverse problem and study its feasibility. Two tuning parameters were selected based on the generalized cross-validation criterion(GCV) and it was implement by the generic agrithorm(GA). An efficient solver was found and the inverse solutions were obtained as the linear combination of the one-parameter regularized solutions. Compared with MNE, the proposed method utilizing two-parameter regularization can get smaller overall mean squared error(MSE), and can reconstruct the shape of the time course of source better.We employed the singular value decomposition(SVD) of the sensor data to gain the temporal characteristics. Specifically, the MEG data was assumed to be linear combinations of the source signals, the singular value decomposition(SVD) of the sensor data was implemented and the the temporal subspace of source was defined by a set of the right singular vector. Then we got a new linear equation by projecting the sensor data and the source signals onto the temporal subspace. The dynamic inverse solutions were obtained by projecting the estimate to the new equation onto the solution space. Compared with the dual-parameter regularization method employing temporal smoothness constraint, the proposed method based on SVD of MEG data was easier to implement and the computation was much less. At different noise levels of the sensor data, this method can get smaller overall MSE and can reconstruct the shape of the time course of source better at a higer signal to noise ratio(SNR).