Hi
On 20 Dec 2006, at 21:15, Stephane Jacobs wrote:
> Hello everyone,
>
> In the process of trying to understand cluster thresholding, I
> played around
> with some different and sometimes non-sense values both for the Z
> score and
> the P value thresholds.
>
Good, but be aware that if you're using non-sense values don't expect
to be able to interpret results ...
> My understanding of the cluster thresholding is that it comprises 2
> steps:
> 1) a Z score threshold is apply to every voxel, keeping only the
> ones that
> pass the threshold.
yep
> 2) the map obtained for the first step is thresholded again using
> the chosen
> P value, which determines a limit size for the clusters to be
> considered
> significant. This size is obtained from the gaussian field theory,
> but it's
> unclear to me how exactly the P value and the size correspond, if
> anybody
> wants to take a shot at explaining this... (this is actually my first
> question here I guess!)
>
Well, the probability of one or more super-thresholded cluster is
approximated by the expected Euler characteristics (number of blobs -
number of holes). This approximation is something like
E(EC) = c * (Z^2-1) * exp (-1/2*Z^2)
where c is some constant that depends on the smoothness, the size of
the search region and Z is your threshold. This approximation only
makes sense when EC is less than 1 and only becomes somewhat accurate
once EC get's close to 0...
If you plot such a curve you'll find that EC is approaching 0 as Z
gets large _or_ as Z gets close to +-1. For high Z the EC basically
is the number of blobs only - our implementation deals nicely with
this case, For low Z you'd also need to include the number of holes.
For Z->0 the EC should be negative. Using the FSL implementation will
not give interpretable results in this range...
>
> So, in one of my thresholding experimentations, I have thresholded
> my data
> to Z = 0 and P = 1.0, which means to me NO THRESHOLDING at all -
> all the
> voxels pass the Z score threshold (0), and the P value allows all the
> clusters as small as a single voxel to be considered significant.
Not really, for reasons outlined above... the expression for p is not
analytic, it's approximated by the expected EC which is not valid in
this range if you only count the number of blobs...
Bottom line: GRF via E(EC) makes sense only at reasonable smoothness
levels and for Z >> 1.
cheers
christian
> This is a
> stupid thresholding, but I just wanted to check that I had understood
> everything properly. Thus, I expected to get a really noisy map,
> with the
> whole brain lighting up (nice for Christmas...). To my wonder, I
> get totally
> reasonable patterns of activations!
> Of importance is that what I describe here happens when I do a
> paired-comparison analysis. When I just contrast each of my
> experimental
> conditions vs. rest, I do get a nice Christmas tree, as expected.
> Could
> anybody tell me the reason for that?
>
> Thanks a lot for any help,
>
> Stephane
--
Christian F. Beckmann
Oxford University Centre for Functional
Magnetic Resonance Imaging of the Brain,
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
Email: [log in to unmask] - http://www.fmrib.ox.ac.uk/~beckmann/
Phone: +44(0)1865 222551 Fax: +44(0)1865 222717
|