Dear Kevin and other M/EEG users,
Following the previous e-mail exchange I got the message below from
Bernhard Spitzer who kindly permitted me to forward it to the list.
Bernhard probably has the most experience with reconstructing induced
responses with SPM so I hope you will find his tips useful. Also note
that this functionality is work in progress. There are many aspects of
it that can be optimized and we rely on your feedback to do that.
Best,
Vladimir
-----------------------
Dear Vladimir,
the recent posts to the mailing list reminded me to share my
progress with imaging induced EEG responses using SPM.
Before the bug-fix I wasn't happy at all with the results (naturally..)
and switched to dics instead, where the results were
after all moderately satisfying. However since the bug-fix,
I clearly prefer the SPM Imaging, it gives much more sparse
solutions than the beamformer, and my general impression is
that everything works as it should.
Empirically, my experimenting fully confirms your recent comments
on how to optimize the procedure (empirically):
-Group inversion works great, and is really comfortable to use.
-Prefiltering the data around the band of interest definitely improves
things, in particular for effects which are subtle on the channel level.
I do it by adding a few lines to spm_eeg_inv_group.m
(inverse.lpf and inverse.hpf).
-Also, restricting the time window to roughly the time of interest
(inverse.woi) can improve the solutions. This may in particular apply
to my data, where I have relatively long epochs of > 5s, with different
induced effects at different times.
I also made progress with our attempts to image power changes
relative to a prestimulus baseline. The procedure that turned out
to work for me is as follows:
E.g., for imaging an induced alpha increase (ERS) at 500 to1000 ms,
relative to a -750 to -250 ms baseline, I use group inversion with the settings
inverse.trials={'211','261','311','361'}; % four conditions, want to
show a parametric ERS effect later
inverse.lpf=7; % prefiltering
around alpha
inverse.hpf=13; inverse.Han=0;
% No hanning, to retain full power both pre- and post stimulus
inverse.woi=[-1000 1250]; % time window long enough
to capture both the pre- and post stimulus epoch
(slightly longer woi, because else the contrast hanning windows will
be cut at the edges)
From the inversion, I output windowed contrast images for 500 to 1000
ms and for -750 to -250 ms,
each four images, 8-12Hz, 'induced'. In spm_eeg_inv_mesh2Voxels.m, I
turn the grand mean
scaling OFF (scale= 1, see below).
I always smooth the images, even with group inversion, et least 8 8 8 FWHM.
I did not yet systematically explore effects of different kernel sizes, though.
I than compute for each of my four conditions the 'relative change'
post- vs pre-stimulus
with ImCalc, either
log(i1+(1/8)) - log(i2+(1/8)) as proposed by Karl
or a standard ERD/ERS 'measure' like
(i1-i2)./i2, which is analogous to the way ERD/ERS is computed in the
channel analysis.
In my experience, both computations give highly similar results in
statistical analysis.
Finally, I do a standard 2nd level analysis on these images. I am
happy with the results!
I also have sometimes a special situation where my data have stimulus artefacts
(0-1250ms) and I want to image ERD/ERS differences at ~2000ms. I than do
two seperate inversions, one for -1000 to 0, and one from 1500-2500.
In that case I found
it advantageous to turn the gm scaling in mesh2voxel back to ON.
Generally I found that gm scaling on/off does not severely change the
location of the sources,
but it can affect the overall magnitude of pre/post stimulus power
changes. E.g. if channel analysis
indicates a massive alpha increase over posterior channels, and a
moderate alpha decrease
over frontal channels, source analysis with gm scaling 'on' may return
only a weak occipital
power increase (positive t contrast) but a massive frontal power
decrease (negative t contrast).
In my experience, under most circumstances, scale=1 seems to better
reflect the size of pre-/post
power changes found on the channel level.
|