MEG Source Detection Revisited

摘要

Magnetoencephalography (MEG) is a multi-channel imaging technique. It uses an array composed of a large number of Superconducting Quantum Interference Device (SQUID) to measure the magnetic fields produced by the primary electric currents inside the brain. The measured spatio-temporal magnetic fields are then used to estimate the locations and strengths of these electric currents, often known as MEG sources. The estimated quantities are finally superimposed with the images generated by magnetic resonance imaging (MRI). The combination of information from MEG and MRI forms the magnetic source image (MSI).A great variety of signal processing and modeling techniques such as Inverse problem, Subspace approach, Independent component analysis (ICA) method, and Beamforming (BF) are used to estimate these sources. The first three approaches require the number of sources be detected a priori. Several shortcomings exist in the currently used methods for detecting the source number. First, the source detection is completed only after - not before - MSI is generated. Secondly, the detection methods are somewhat subjective.In order to provide a solution to the problem of detecting MEG source number for all these approaches, a novel method is developed. The covariance matrix of MEG measurements over all channels is decomposed into the signal and the noise subspaces. The number of sources is shown to be equal to the dimension of the signal subspace. The selection of this dimension is translated into a problem of determining the order of the underlying statistics. This statistical identification is resolved by using Information theoretic criteria which are derived based on Kullback-Leibler divergence. Because the method utilizes originally acquired MEG measurements and implemented before magnetic source images are generated, it is an entirely data-driven approach, more efficient, and less likely to be subjective.