<div dir="ltr">Dear Eelke, very interesting topic. Please find my two cents below:<div class="gmail_extra"><br><br><div class="gmail_quote">On Thu, Nov 21, 2013 at 11:36 AM, Eelke Spaak <span dir="ltr"><<a href="mailto:eelke.spaak@donders.ru.nl" target="_blank">eelke.spaak@donders.ru.nl</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Fellow FieldTrippers,<br>
<br>
Currently I am looking at a contrast for left- versus right-hand index<br>
finger button presses. As expected, on sensor level (combined planar<br>
gradient, grand average) I see a clear lateralisation in beta band<br>
power starting at least 0.5s before the button press (see<br>
<a href="https://db.tt/Rtch3Qjy" target="_blank">https://db.tt/Rtch3Qjy</a>). Both 'blobs' are significant; there is<br>
clearly more beta power ipsilateral to the response hand. I would<br>
prefer to do further analyses on source level, so I attempt to<br>
reconstruct the sources for this effect using DICS beamformer (common<br>
filter, applied to both conditions separately; fixedori and realfilter<br>
= 'yes'). The grand average results for this (again contrast left vs<br>
right response hand) are shown at <a href="https://db.tt/IBQZG0d8" target="_blank">https://db.tt/IBQZG0d8</a> . (Ignore the<br>
R/L-flip, this is radiological convention.)<br></blockquote><div><br></div><div><br></div><div>There might be issues with:</div><div>- head positions/distance of single subjects in the dewar</div><div>- something happening in the transition axial -> planar gradiometers conversion</div>
<div>- orientation of the brain motor sources</div><div>- inverse problem reconstruction</div><div>- magnitude of the power effects for the different subjects</div><div>- forward model issues </div><div><br></div><div>Let me comment briefly on point 4. If the sources in the motor cortex are bilateral (as expected to different extents from ipsi to contra) and are temporally correlated, this constitutes an issue for the beamformer algorithm (van Veen et al.1993) especially if the sources are near (and lead fields highly correlated).</div>
<div>There are workarounds to localize the single contributions of primary/secondary motor sources, but this implies the use of regional suppression (a nulling beamformer) and it is a tedious procedure to apply. Might be worth to look into that though. Let me know if you have interest in this.</div>
<div><br></div><div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
As you can see, the source level solution is much more blurry than on<br>
sensor level. This picture is without using any regularisation (lambda<br>
parameter), the results are even worse when I use lambda = '5%'. The<br>
negative blob (right hand higher power than left) becomes 'marginally<br>
significant' on source level (p ~ 0.06) where it was p < 0.001 on<br>
sensor level. The positive blob is nowhere near significant. Also, the<br>
individual results are much less topographically consistent on source<br>
than on sensor level (explaining the worse statistics).<br></blockquote><div><br></div><div>On this point: using regularization is useful to invert the real(csd) matrix if it is ill-conditioned, but might blur the source reconstruction. On the other hand not using it might also be detrimental on the source reconstruction. Therefore applying it depends on the condition number of your csd. Is the matrix full rank? If not you might consider regularizing.</div>
<div>Did you previously use ICA/PCA to get rid of artifacts? If yes that will have a negative effect on the condition number. and you might consider cranking down the ICA components rejection to the big spiky components -if any- and let the beamformer filter reject the smaller ones (in magnitude).</div>
<div>Are you using cluster statistics based on the maxval? Maybe another statistics might give different results, given that the maxval might be due to artifacts (muscular, heart, ...)</div><div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
I have checked the segmentation of my MRIs, the 'gray' seems to be<br>
nicely within the head all the time. Also, I have manually verified<br>
the alignment of headmodel, sourcemodel, and gradiometer information<br>
for all subjects.<br></blockquote><div><br></div><div>Do all subjects sit with the top of the head at the same distance from the dewar? This is easily visualized by plotting the headmodel together with the head coordinates' sensors.</div>
<div>By experience with CTF systems not all subjects' heads are at the same distance from the top (because they slip down, or they reposition 'cause they can't see out of the dewar, they also move!). This might hinder the SNR of the raw data to start with.</div>
<div>It seems the grand average is still significant, but might be due to the contribution of a few good subjects.</div><div>You may also want to consider the movement correction GLM method at the source level. Maybe it gives you back some SNR.</div>
<div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
As a final note, the above sensor-level plot was taken from a 'slice'<br>
out of a planar-gradient time-frequency analysis (mtmconvol). The<br>
ingredient for the beamformer was an mtmfft fourier spectrum on the<br>
axial gradiometer data, obtained for just the time-frequency range of<br>
interest (subselect toilim [-0.5 0], mtmfft foi = 23, tapsmofrq = 7).<br>
When I compute condition-averaged power based on these fourier spectra<br>
and look at the contrast, the results are again as expected:<br>
<a href="https://db.tt/n2P3UKcQ" target="_blank">https://db.tt/n2P3UKcQ</a> (of course less localised because of axial<br>
gradient vs planar). The freq structures underlying this contrast are<br>
exactly the same as those going into ft_sourceanalysis, so the problem<br>
must be in the source analysis step (and/or in the preparation of the<br>
geometric information, although these seem fine by visual inspection).<br>
<br>
Does anyone have any idea that might explain these seemingly<br>
contradictory results? I would have expected demixing to improve<br>
signal-to-noise ratio, rather than worsen it.<br></blockquote><div><br></div><div>Yes, but it can also depend on different extents from FFT-level analysis. Have you played around with spectral smoothing, by for example lowering the number of tapers, or changing to a single condition contrast like activation vs baseline in the beamformer?</div>
<div>What is the grand average vs single subject lateralization effect/single condition effect/ planar or axial effect? Is it there for all single subjects?</div><div><br></div><div>I hope the comments helped a bit.</div>
<div><br></div><div>All the best!</div><div>Cristiano</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
Thanks!<br>
Best,<br>
Eelke<br>
_______________________________________________<br>
fieldtrip mailing list<br>
<a href="mailto:fieldtrip@donders.ru.nl">fieldtrip@donders.ru.nl</a><br>
<a href="http://mailman.science.ru.nl/mailman/listinfo/fieldtrip" target="_blank">http://mailman.science.ru.nl/mailman/listinfo/fieldtrip</a><br>
</blockquote></div><br></div></div>