r.oostenveld at FCDONDERS.RU.NL
Sat Aug 20 10:08:19 CEST 2005
First some general comments.
In my experience megrealign works more accurate with the sincos
option (i.e. nearest neighbour interpolation) than with the
sourceproject option. I have validated this function using simulated
data, in which I computed the field of a dipole on axial
gradiometers, and the field of the same data on (true) planar
gradiometers. Then I converted the axial to the planar data with
megplanar and the different options there and compared the two. I
repeated this with different noise levels and the sincos was
consistenly more acurate than the others. However (relevant for other
people as well) the 'orig' method is still the default. That was
implemented initially by Ole Jensen, the other ones were implemented
The low-level forward computation of the fields in Fieldtrip cannot
make use of the synthetic higfher order gradients. In theory that
could be supported, but the weights of the reference sensors are not
read in from teh res4 file. That means, whenever you do a computation
that involves a forward computed field (sourceanalysis,
dipolefitting, megrealign and megplanar+sourceproject), the forward
field is a plain 1st order MEG gradient. If your CTFdata is 3rd
order, and you have saved it with newDs, there will be a slight
discrepancy between the measured data and the forward model, even if
your measured data contains a single perfect dipole. The field of the
sources in the brain will also be slightly supporessed, which is not
represented in the forward model. However, the advantage of the 3rd
order gradient is that also environmental noise is suppressed. So
there is a trade off, if you have a lot of noise, it is better to use
the 3rd order gradient, if you have very clean data, it is better not
to use it.
I would like to implement the 3rd order gradients, and have most code
for it in place, but there are some details that need to be sorted
out. If you are going to use it and think that you can contribute, I
would appreciate your help on that.
Now onto your question:
On 19-aug-2005, at 23:43, Sanja Kovacevic wrote:
> Oh, I forgot to mention that I used meplanar with the source
> projection option. I tracked the problem down to axial2planar (line
> 309 of megplanar: planar.grad = axial2planar(, axial.grad)).
> Axial2planar does not return the field "type" in planar.grad, and
> therefore compute_leadfield could not recognize the sensor type and
> gives "the forward model does not look like EEG, nor like MEG" error.
One potential problem here might be that axial2planar is also trying
to compute the planar representation of the reference sensors. I
don't think that that applies in your case, but it should not do
that. So that is something to check.
The latest version of compute_leadfield (revision 1.7) should also
detect the correct type in the absence of the "type" field. It has
the following added with repect to the older versions
elseif isfield(sens, 'pnt') & isfield(sens, 'ori')
iseeg = 0;
ismeg = 1;
If yours does not contain that, please upgrade your fieldtrip version.
More information about the fieldtrip