Optimization with the living brain on a millisecond time scale using MEG
Speaker: Alexandre Gramfort , INRIA, France / Harvard Medical School
Date: February 10 2011
Time: 11:30AM to 12:30PM
Location: 32-D507
Host: Polina Golland, CSAIL
Contact: Polina Golland, x38005, polina@mit.edu
Relevant URL: MEG offers the possibility to measure the magnetic field
induced by brain activity on a millisecond time scale. The
mathematical and computational challenge is
then to localize in space and in time the active brain regions
at the origin of the signal measured. Due to the physical
properties of the head and the limited number of sensors
the problem is ill-posed. Priors are therefore required
to obtain solutions.
While the majority of the solvers used in the MEG community rely
on convenient L2 priors (Tikhonov), many contributions
since the 90's have proposed alternative priors whose main
goal is the improvement of the spatial resolution of brain
imaging with MEG. A recurrent issue with such alternative
solvers is their computation time usually incompatible with
real case studies.
Borrowing ideas and results in both signal processing and
machine learning communities, I will detail how convex
sparsity inducing priors can be used to improve source
localization results with MEG while keeping reasonable
computation times.
I will present how 2-level (see [Ou et al. Neuroimage 2009])
and 3-level mixed-norms can be used in particular
when computing the inverse solution for multiple datasets
jointly. I will then explain how sparse priors
in Gabor dictionaries can used to improve the temporal
estimates of source activations.
The optimization procedures used are based on first order
methods and proximal operators. When possible
computation times are significantly improved with
active-set strategies.
See other events that are part of Biomedical Imaging and Analysis 2010/2011
See other events happening in February 2011
