An inversion method based on random sampling for real-time MEG neuroimaging

Abstract

The MagnetoEncephaloGraphy (MEG) has gained great interest in neurorehabilitation training due to its high temporal resolution. The challenge is to localize the active regions of the brain in a fast and accurate way. In this paper we use an inversion method based on random spatial sampling to solve the real-time MEG inverse problem. Several numerical tests on synthetic but realistic data show that the method takes just a few hundredths of a second on a laptop to produce an accurate map of the electric activity inside the brain. Moreover, it requires very little memory storage. For this reasons the random sampling method is particularly attractive in real-time MEG applications.

Less Is Enough: Assessment of the Random Sampling Method for the Analysis of Magnetoencephalography (MEG) Data

“…We point out that although the average of 10 runs is poor from a statistical viewpoint, it is a good compromise between accuracy and computational cost. In [20] the authors investigated how the increasing of the number of runs up to 50 affects the localization accuracy. They found that accuracy does not improve significantly as the number of runs increases.…”

“…Another beamforming method, Truncated Singular Value Decomposition Beamformig (TSBF), presented in [30] was proved to give good results when used for the solution of the MEG inverse problem (cf. [20]). For all the methods, the computation of the inverse solution is given by J = Q B where Q is the inverse kernel of the method.…”

“…The RAndoM Sampling mEThod (RAMSET) [19,20] meets these demands. In fact, RAMSET uses a sampling procedure to significantly reduce the dimension of the source space thus reducing both memory usage and computation time.…”

“…In fact, RAMSET uses a sampling procedure to significantly reduce the dimension of the source space thus reducing both memory usage and computation time. A first study on the impact of this source space reduction on the estimate of the brain activity generated by a single neural source was presented in [20] using synthetic MEG data. In this paper we test RAMSET on both synthetic data and real MEG measurements and we investigate how the source space reduction performed by RAMSET affects the localization accuracy of well established inverse algorithms when the forward problem is solved by three different methods.…”

“…The matrix L is regularized by means of a truncated singular value decomposition, employing 80 singular values, and τ is set to 10 −8 (cf. [20]).…”