Dear Kalina and Ivan,
first regarding Ivans question.
The "original" results for Gaussian Random Fields assume that volumes where
"large compared to the smoothness", at least I think that was the formulation,
meaning basically that the effects of the edges could be disregarded. Worsley et
al. later came up with generalisations which allowed the search volume to take
any shape. However, the threshold will be a function of the smoothness, volume
and surface of the search volume. For a given volume the lowest threshold (and
highest sensitivity) will be given by the shape which minimises the surface
(i.e. a sphere).
Hence, you are right that using a grey matter mask may not increase sensitivity
(I don't think it would) but to my undestanding this should not be due to
changes in smoothness (nor to changes in its estimate) but rather to a change in
the volume to surface configuration.
Having said that, it is likely the smoothness estimate will be affected by the
inclusion/exclusion of the ventricles since these often contain "unmodelled
effects" whic affects the estimate of the smoothness based on the residuals.
Secondly regarding Kalinas suggestion
The AND across the time series performed by SPM on the thresholded images do
tend to erode the brain, and as you have noticed in particular in areas with a
high susceptibility gradient. However, this is not just a bad thing (and perhaps
not even predominantly bad) since it offers some level of protection from false
positives due to movement related variance in areas which are extremly sensitive
to this. The movement related variance does not have to be much correlated to
the task, to show up as task effects in these areas.
Hence, if one is to skip the mask I would suggest (at least) including the
estimated movement parameters and the estimated movement parameters squared in
the design matrix. In addition, any task related swallowing or tounge movements
would be a problem.
Jesper
> Dear Ivan,
>
> To my understanding (and from some limited experience with this), using a
> mask image at the model estimation stage to reduce the search volume does
> not affect the uncorrected Z-score estimate - the uncorrected p-values are
> independently estimated for each voxel, and so they won't improve (or
> change in any way) by changing the mask. The corrected p-values, on the
> other hand, would benefit from reducing the search volume (although other
> factors such as change in smoothness would also be relevant).
>
> Regarding the practical side of it: perhaps one possibility to ensure the
> mask doesn't have sharp edges (so that it doesn't decrease the smoothness
> estimate) is to simply smooth it with a large kernel at some stage of its
> creation.
>
> In order to explicitly specify a mask at the individual model estimation
> stage, you need to change several things in the SPMcfg.mat file
> before esimation (this was described on the list some time ago). I
> recently wrote a small program to do this automatically - it can be
> downloaded from:
>
> http://www-psych.stanford.edu/~kalina/SPM99/Tools/glm_specmask.html
>
> The reason for writing this program was that we've noticed here that
> on many occasions, the default mask (based on 80% of the global
> signal) would not include regions in or near susceptibility areas -
> especially for functional images collected at higher fields (e.g.,
> 3T) where differences in intensity between regions appear the be
> larger. On the other hand, when an explicit mask based on the segmented
> gray matter is used during model estimation, often times there appear to
> be robust effects in regions that would have been excluded otherwise
> because of low signal values.
>
> I was wondering if anyone else has any similar observations and what other
> approaches exist for dealing with this. As it appears to be potentially an
> important issue (it may lead to claims about the absence of activation in
> regions that were activated by the paradigm), I was wondering to what
> extent it has been identified as an issue by other groups as well. Any
> comments would be much appreciated.
>
> Best regards,
> Kalina
>
> On Thu, 7 Dec 2000, Ivan Toni wrote:
>
> > Dear all,
> >
> > I'm trying to use an anatomical mask on an EPI timeseries. Previous
> > messages in this list have addressed a related topic, i.e. the use of an
> > anatomica mask at the time of contrast estimation. However, I'm trying
> > to get a bit further than using a mask just for cosmetic reasons. My
> > goal is to try to constrain the volume assessed by the statistical
> > procedure, so that hopefully this a-priori information might improve the
> > statistical power. More precisely, I thought that masking a dataset with
> > a proper anatomical mask (let's say, a smoothed segmented gm), should
> > reduced the search volume and therefore the Z score might improve.
> > However, I don't know what is going to happen to the smoothness of the
> > image. If the smoothness of the residuals is reduced, then the
> > statistics would actually get worst.
> >
> > My first question is the following: does anybody have any experience on
> > this topic, i.e. if you reduce the search volume to gray
> > matter with an anatomical mask (rather than to the usual 80% of global
> > signal), does this have a positive effect on the stat ? I vaguely
> > remember a paper of Woolsey where he claimed that sharp edges in a mask
> > image might actually reduce the image smoothness, with a negative effect
> > on the stat.
> >
> > Second question: provided it is worth to implement this anatomical mask
> > at the time of model assessment, how do you tell SPM to do it ? I saw
> > there is a field xM, in spm_fmri_spm_ui, with a flag xM.I set to zero,
> > so that SPM is extracting its own mask from the actual EPI timeseries.
> > However, when I put this xM.I = 1, and the xM.VM = {['mymask.img']},
> > it's not very happy (I got an error like "improper index matrix
> > reference" in spm_spm line 689).
> > I also tried to induce SPM to mask the dataset by masking just one EPI
> > image of the timeseries with my anatomical mask. Since SPM is
> > considering only voxels common to the whole dataset, this should have
> > worked. However, in this case I got an error message at the time of
> > model estimation ( Error in spm_spm.m on line 689 ==> tM =
> > inv(xM.VM(i).mat)*M; %-Reorientation matrix
> > Error in spm_fmri_spm_ui.m on line 622 ==>
> > spm_spm(VY,xX,xM,F_iX0,Sess,xsDes);
> >
> > Thanks in advance,
> > Yours,
> > Ivan Toni
> >
> >
>
> _____________________________________________________________________________
> Kalina Christoff Email: [log in to unmask]
> Office: Rm.478; (650) 725-0797
> Department of Psychology Home: (408) 245-2579
> Jordan Hall, Main Quad Fax: (650) 725-5699
> Stanford, CA 94305-2130 http://www-psych.stanford.edu/~kalina/
> _____________________________________________________________________________
|