Magnetoencephalography (MEG) is a multi-channel, functional imaging technique. It measures the magnetic
field produced by the primary electric currents inside the brain via a sensor array composed of a large number
of superconducting quantum interference devices. The measurements are then used to estimate the locations,
strengths, and orientations of these electric currents. This magnetic source imaging technique encompasses
a great variety of signal processing and modeling techniques which include Inverse problem, MUltiple SIgnal
Classification (MUSIC), Beamforming (BF), and Independent Component Analysis (ICA) method.
A key problem with Inverse problem, MUSIC and ICA methods is that the number of sources must be
detected a priori. Although BF method scans the source space on a point-to-point basis, the selection of
peaks as sources, however, is finally made by subjective thresholding. In practice expert data analysts often
select results based on physiological plausibility. This paper presents an eigenstructure approach for the
source number detection in MEG neuroimaging. By sorting eigenvalues of the estimated covariance matrix
of the acquired MEG data, the measured data space is partitioned into the signal and noise subspaces. The
partition is implemented by utilizing information theoretic criteria. The order of the signal subspace gives
an estimate of the number of sources. The approach does not refer to any model or hypothesis, hence, is an
entirely data-led operation. It possesses clear physical interpretation and efficient computation procedure.
The theoretical derivation of this method and the results obtained by using the real MEG data are included
to demonstrates their agreement and the promise of the proposed approach.