[pymvpa] Preprocessing in FSL, question on specifics

[Susanne Greiner]

Dear Mike,

first of all I am not a neuroscientist but more of a machine learner 
(who has had your problem a lot of times).
I know who does just as you describe and just concatenates first and 
applies motion correction for the whole session. Sessions are usually 
never co-registered just by motion correction even if it is possible.
There are also others who perform motion correction on each run, and 
take one run/session as reference and co-register all others with it and 
furthermore co-register between sessions.

Obviously the better the features/voxels match between runs or sessions, 
the better any kind of prediction can be expected to work.
For this reason consider:
Whenever you do some kind of motion correction or co-registration you 
interpolate somehow and introduce some kind of smoothing. This means if 
you do multiple actions like that it might happen that you introduce a 
notable amount of smoothing to all runs which have not been used as 
reference.
For this reason there are also 'crazy ones' that take one volume/run 
(preferably with the higher contrast so within the very first volumes of 
the run which you usually discard later) and then you select a reference 
and co-register the other to it. The mean of these co-registered HC 
columns is taken as reference for motion correction for all runs, so 
every run has 'almost' the same amount of smoothing-interpolation added 
to the original data (this has been never demonstrated though).

As you see, there are quite a few ways of doing it (maybe others I have 
never come across yet). As far as I know there are few/ no publications 
out there stating exactly what works best when... (just every lab has 
its way of doing it)
I would recommend you to use the method, that makes you feel most 
comfortable (and shows best voxel correspondence when having a short 
look at what you did)

Note that between runs/sessions there might also be important 
differences which need z-scoring and linear de-trending (easily done by 
pymvpa not fsl just before using the classifier)

If you discover your own way of solving the problem and it works out 
well, I encourage you to share the knowledge!!

LG,

Susanne

PS: I just saw, Michael has already answered. As he states, never think 
that you will be able to solve everything by pre-processing :-)

--------------
> Hi all,
>
> I've used FSL/FEAT before, though never specifically for preprocessing 
> for PyMVPA, so please forgive the simplicity of these questions. I 
> have an experimental design with several short sessions, so for a 
> standard FSL analysis (i.e. univariate) I've performed motion 
> correction (etc.) and stats on each session separately, before running 
> a second level analysis on the FEAT results. (I should also note that 
> our scanner will output a separate 4D nifti file for each short session.)
>
> So essentially I'm wondering the best order of events: (a) when to 
> concatenate the shorter 4D files into one large 4D file, (b) whether I 
> should run motion correction 1 time (after the concatenation) or 
> whether it should be run separately for each session's nifti file, 
> before being run a second time on the concatenated file. While there 
> is very little head motion within each session, there looks to be 
> considerably more between sessions, which probably comes as no 
> surprise. A test of running motion correction a single time (after 
> concatenation) looks like it does not perform very well: there is 
> still a large amount of motion visible to the naked eye.
>
> As an aside, if I plan to discard some specific volumes, I also would 
> value any input as to whether it makes more sense to delete them from 
> the 4D time series (using fslsplit and fslmerge), or to leave them in 
> and give them their own label ("discard") in the attributes.txt file.
>
> Best and many thanks,
> Mike
>
>
> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20110712/adf2d451/attachment-0001.html>