Hi Darren,
>I am getting clusters that contain 20,000
> > voxels. For instance:
> >
> > Cluster List
> >
> > Cluster Index Voxels P -log10(P) Max Z x (mm)
> y (mm) z (mm)
> >
> 2 20330 0 48.6 8.06
> -30 -92 4
> > 1 675 0.000699
> 3.16 5.23 26 20 26
Yep, that's not too surprising. The thresholding defaults in the FEAT
analysis aren't very helpful (no offense Steve -- I don't think any default
would ever be generally helpful across fMRI studies). Setting the
thresholds appropriately involves some effort.
First, it's important to be clear about what question(s) you want to ask of
your data. For initial analyses, I typically just want to get a feel for
what activity is present in my contrasts. To do this, I generally use
voxel level statistics (rather than clusters) because I don't care at this
stage how big my clusters are. I just want to know where the activation
is. In Feat4 there was a bug in voxel-stats, though, such that it didn't
produce the table of results so I typically use cluster stats with a
Z-value of either 2.3 or 3.1 (depending on how lenient I want to be) and a
p value of 1.0 -- which is equivalent to an uncorrected voxel level p-value
of 0.001 for Z=3.1.
The issue in your data seems to come from the nature of a cluster level
statistic. Briefly, this is based on setting a completely arbitrary height
threshold (the Z value, 2.3 is the default in FEAT) and then only accepting
clusters above a given size (extent) as significant. THe size is
determined by the height threshold you chose plus the volume of data and
its smoothness. Essentially, what you wiill accept as significant is any
cluster which is larger than you would expect *by chance* when thresholding
a gaussian random field at the specified height threshold. Obviously the
chance clusters will be larger the lower the height threshold. For a low
height threshold such as 2.3, you'd need HUGE clusters for them to be
significant. Thus your 20,000 voxel cluster in one case and no clusters in
the other. But as you said, this is hardly meaningful. What you probably
want is to use a higher Z-value (maybe 3.1, maybe higher) and possible
loosen up the p-value a bit -- especially for data exploration. If you are
looking for activation in a specific area, cluster stats may be the wrong
way to go. Matthew Brett has an excellent web page with some sample matlab
code to illustrate voxel and cluster stats
(www.mrc-cbu.cam.ac.uk/Imaging/randomfields.html).
In my experience, it's well worth working through the differences between
voxel and cluster statistics (at least conceptually) in some detail because
it is crucial to asking the questions one's interested in. I think this is
one of the single biggest difficulties in using cluster stats for most
people but once you crack it, it's easy and you'll never have trouble with
it again.
Hope this was some help.
Joe
--------------------
Joseph T. Devlin, Ph. D.
FMRIB Centre, Dept. of Clinical Neurology
University of Oxford
John Radcliffe Hospital
Headley Way, Headington
Oxford OX3 9DU
Phone: 01865 222 738
Email: [log in to unmask]
|