Hi - I believe that this is correct, yes.
Cheers, Steve.
On 27 Mar 2008, at 11:07, Renate Schweizer wrote:
> Dear Tom,
>
> thanks a lot for all the helpful information.
> I have one related question, just to make sure that I got everything
> right. I guess the 3-voxel FWHM rule of thumb does not mean, that I
> have to smooth with a "3-times-voxel-size" mm Gaussian kernel, but
> is the smoothness of the data.
> I know that SPM gives this value as "Smoothness FWHM (mm: x,y,z),
> (voxel: x,y,z)", the later being multiplied to calculate the number
> of voxels per resel. In FSL I found so far only "RESELS 3.8" which I
> interpret as 3.8 voxels per resel. If I take the 3.8 to the power
> of 1/3 (= 1,56) would that be the voxel FWHM ? Or does FSL give the
> FWHM in mm somewhere ?
>
> Greetings,
>
> Renate
>
>
>
>> Dear Renate,
>>
>> The rule of thumb is that 3-voxel FWHM is generally suffice to
>> ensure accurate RFT results *when* there is sufficient degrees-of-
>> freedom, like 50 or more (as you probably have since you're doing
>> single subject fMRI?). For low DF, much more smoothness is needed,
>> so much so you're probably just better off using randomise.
>>
>> For voxel-wise inference, when the smoothness is too low, the
>> results become increasing conservative (in fact, Bonferroni can be
>> less conservative, but I'm pretty sure that FSL checks and uses the
>> better of Bonferroni and RFT). For cluster-wise inference RFT can
>> become very unstable, being either quite conservative or anti-
>> conservative (liberal).
>>
>> One way around this problem is to use FDR, which doesn't require
>> smoothness to be valid.
>>
>> -Tom
>>
>> Refs:
>>
>> Voxel-wise RFT:
>> T.E. Nichols and S. Hayasaka. Controlling the Familywise Error Rate
>> in Functional Neuroimaging: A Comparative Review. Statistical
>> Methods in Medical Research, 12:419-446, 2003.
>>
>> Cluster-wise RFT:
>> S. Hayasaka and T.E. Nichols. Validating cluster size inference:
>> random field and permutation methods. NeuroImage, 20:2343-2356, 2003.
>>
>>
>> I have high resolution, small volume fMRI data (1mmx1mmx2mm, 3
>> slices) on which I want to do only limited spatial smoothing, if at
>> all, since I want to retain the resolution and the expected
>> activations are rather small. For a first approach I smoothed with a
>> Gaussian kernel of 1mm FWHM and thresholded applying a ROI mask and a
>> cluster threshold at Z=2.3, p=0.05.
>> I got reasonable results, but my concern is that these results may
>> not be valid because of the limited smoothness of the data. The
>> calculated smoothness is minimal (DLH 1.3, VOLUME 21520, RESELS 3.8)
>> possibly violating the assumtions of the GRF Theory which expects a
>> "certain smoothness". So far I haven't found any descriptions of
>> numbers for that required smoothness....
>> Are there any parameters, e.g. in the results of the thresholding,
>> which would indicate that the assumptions concerning the smoothness
>> are violated ?
>>
>> Greetings and thanks for the help,
>>
>> Renate
>>
>> ____________________________________________
>> Thomas Nichols, PhD
>> Director, Modelling & Genetics
>> GlaxoSmithKline Clinical Imaging Centre
>>
>> Senior Research Fellow
>> Oxford University FMRIB Centre
>
>
>
---------------------------------------------------------------------------
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director, Oxford University FMRIB Centre
FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~steve
---------------------------------------------------------------------------
|