[FieldTrip] source_recontruction with intracranial sensors

mcpiastra at libero.it mcpiastra at libero.it
Fri Jul 5 17:47:32 CEST 2013

Dear Fieldtrip developers and users,

I have some problems with source reconstruction starting with intracranial 
sensors. Schematic details are below.

Aim: do source reconstruction using BEM for the forward problem and mne for 
the inverse problem.
Dataset: stereoEEG recording (50 sec)
Notation: I'll use the same nomenclature of the one in the tutorial

Since we have intracranial sensors, our aim is to build the forward model 
(leadfield structure) starting from a mesh whose cortical sheet has a higher 
spatial resolution. To do that we approached the problem in two ways:
1) we substituted the innermost sheet, built by ft_prepare_mesh, with the 
Freesurfer cortical mesh (4180 vertices). We put dipoles on the mesh nodes and 
then we built the head model and calculate the leadfield matrices (one for each 
dipole). Solving the inverse problem, in source.avg.pow there were values with 
order of magnitude of 10^{-30}- 10^{-20}. This is due to the fact that the 
matrix vol.mat (one of the output of ft_prepare_headmodel) had values very 
different if we compared the entries in correspondence to the Fieldtrip meshes 
and the ones in correspondence to the Freesurfer mesh. This difference had the 
order of magnitude equal to 10^{9}.

 Due to this pathology, we tried a different way to compute the forward model 
(following more the tutorial).

	2) we built the head model using the default Fieldtrip meshes (all of the 
three sheets) without substituting any sheet. (We just fixed the number of 
nodes for the innermost sheet equal to the Freesurfer one.) This avoided high 
differences in vol.mat. 
Freesurfer mesh vertices were used to determine dipole positions, as we did 
before. In this case we obtained values of source.avg.pow of the order of 10^

In both cases we preprocessed the dataset with ft_timelockanalysis and we used 
the regularization parameter lambda (for the mne method) equal to 0.2. 

Our questions are:
1) what is vol.mat? Which order of magnitude should have its values? We didn't 
manage to explore the building process of it. (We suppose this matrix has 
electrical information interpolated through the mesh vertices.)
As consequences:
2) which is the order of magnitude expected for the values of source.avg.pow ? 
And which is its  unit of measure?
3) Is it correct the interpretation of: source.avg.mom and source.avg.pow as 
the evolution through time of the electric dipole moment and the radiated power 
of the dipole, respectively?
4) In order to apply a regularization approach to solve the inverse problem we 
though that the best way was to introduce the noise covariation matrix built by 
the function ft_timelockanalysis which is subsequently controlled by the lambda 
parameter. Is there any criterion to choose the best lambda? Can we avoid to 
pass through ft_timelockanalysis to regularize the solution?

Many thanks,

Maria Carla Piastra
PhD student in Bioengeneering c/o BioLab @ DIBRIS 
University of Genova,

More information about the fieldtrip mailing list