Title: Bayesian hierarchical mixture models for MEG brain imaging. 

Abstract: MEG (Magnetoencephalography) is a totally non-invasive
imaging modality, in which the electric activity of the brain is
estimated from the recordings of the weak magnetic fields induced by
it outside the head. This modality has a great potential of
complementing the more traditional EEG (electroencephalogrphy)
measurements, e.g., in the localization of the foci of the onset of
epileptic seizures. The MEG inverse problem is severely ill-posed, and
an additional difficulty is due to the fact that the brain itself is
noisy, so separating the signal of interest from the background signal
is non-trivial. In this talk, some ideas based on Bayesian models are
presented. Instead of trying to filter the useful signal from the
data, a mixture prior model for the source itself is defined, and a
part of this mixture model is constructed so that it is sensitive to
focal sources. It is demonstrated by computed examples that monitoring
the energy partitioning between the different components of the
source, the method could potentially be used even for identifying deep
sources.