Hello, Michal
Here is another solution. I'd be interested to hear others' opinions on
this approach:
Typically, if one of the subjects is missing a voxel (ie - there is NaN
value or a zero in that voxel after normalization), then the RFX
analysis returns a NaN at that voxel. But, if you want to do a simple
random-effects analysis T-test across subjects, you might try to do a
t-test of the parameter estimates using the voxels that are present for
each subject. This means that each voxel will have different degrees
of freedom. So, if you keep track of each voxel's degrees of freedom,
you can still calculate T maps, Z maps and p-values. (note that the
normalization has to be good enough at warping things into place even
if there are missing slices).
Pro: I think that using a dilation would essentially extrapolate data
to where there is none, and that can be risky, so this approach is
safer.
Con: this simple t-test is not as sophisticated and accurate as FLAME.
If this interests you, I can give you some matlab code to implement it.
I'll have to dig into my old code first ...
I hope this is helpful,
-Luis
--
( http://www.eecs.umich.edu/~hernan )
On Jun 25, 2004, at 11:38 AM, Michal Kuniecki wrote:
> Hello,
>
> I decided to post my question again since I haven’t got any answer to
> my
> original posting. If I am asking for something too obvious just tell me
> where is it in the help files or in FSL web course. Thanks in advance.
> Below
> is my unanswered question.
>
>
> Hello Everybody,
>
> To start with the question I would like firstly to say that I have
> similar
> problem to that described by Ed Vessel from U. of Southern California.
> Namely I have only a limited number of slices taken from each subject.
> For
> that reason registration is quite poor and in the result I loose data
> from
> outermost slices because data from one or two subjects are missing.
> This is
> very frustrating and either SPM nor FSL have routines to calculate
> valid
> statistics from not complete set of data. Nevertheless Stephen Smith
> proposed partial solution to this problem – that is to run dilate
> command
> (avwmaths <input> -dil -dil <input>) on the images in reg_standard
> directory. Now my problem is that higher level analysis firstly
> creates such
> directories but afterwards it deletes them – so they appear to be
> temporary.
> It is feasible to create reg_standard directories manually of course
> but the
> problem stays the same, since higher level analysis will delete those
> directories anyway. Hence the questions:
> 1) Is there any way of sparing reg_standard directories in higher level
> analysis ?
> 2) If I create reg_standard directories manually will higher level
> analysis
> use them or it will create its own while deleting those created
> manually ?
>
> With best regards
>
> Michal Kuniecki Ph.D.
> Department of Psychophysiology
> Institute of Psychology
> Jagiellonian University
> Krakow, Poland
>
>
|