Hi,
On Tue, 10 Apr 2012 14:03:31 +0200, Marko Wilke <[log in to unmask]> wrote:
>Hello,
>
>I would like to add two remarks to John's mail:
>
>>> 3) Within Normalise (Write) Writing Options, you must state the Voxel Sizes. Our acquisition voxel sizes are: 1.7 x 1.7 x 3. Would we then state 3 3 3 for the Voxel Size?
>> Smaller voxel sizes are theoretically better (because the random field
>> assumptions are best satisfied), but if the voxels are too small, then
>> you quickly run out of disk space. Generally, I would suggest using
>> the default settings of 2mm isotropic, which is what most people would
>> do.
>
>I think this is an interesting point as I usually normalize my
>functional images (acquired at 3x3x3mm) to 3x3x3mm in order to avoid
>interpolating too much (keep spm from "inventing voxels", so to say :) I
>realize that even that is not quite correct as the scaling to MNI is
>already on the order of 1.4, but thought it is better than to upscale to
>a higher-than-original resolution. The multiple comparison issue also
>comes into this equation at some point as fewer voxels = fewer tests.
>The random field argument is not that much of an issue if you use
>voxelwise FDR, which is independent of smoothness estimates. I must
>admit, though, that I have not systematically investigated this. If
>anyone has a more informed opinion than me on this issue, I would much
>appreciate hearing it.
>
There is an interesting paper of Tom Nichols where the relation between degrees of freedom (df), smoothing and voxel size and the validity of Gaussian Random Field (GRF) theory was investigated:
Thomas Nichols and Satoru Hayasaka
Stat Methods Med Res 2003 12: 41
DOI: 10.1191/0962280203sm341ra
One important message is that not the absolue smoothing size is important, but the ratio between smoothing and voxel size (this is also related to that commonly known rule of thumb to smooth data with at least 2-3 times voxel size). This is simply because a gaussian curve cannot be assumed if that curve is sampled by only a few points. This is even more important for lower df's. Figure 4 in the paper demonstrates that issue. For lower df's and a small ratio between smoothing and voxel size the GRF is deviating more for the expected threshold.
For VBM this means that you are usually on the safe side due to small voxel size of normalized data (1.5mm). For fMRI the voxel size to write normalized data depends on the intended smoothing size. If you smooth your data with FWHM of >=9mm a voxel size of <=3mm is fine. If you intend to apply a smaller FWHM of <9mm I would recommend to write normalized data with 2mm or smaller. With the default value of 2mm you are usually on the safe side and this value is a good tradeoff.
Regards,
Christian
____________________________________________________________________________
Christian Gaser, Ph.D.
Departments of Psychiatry and Neurology
Friedrich-Schiller-University of Jena
Jahnstrasse 3, D-07743 Jena, Germany
Tel: ++49-3641-934752 Fax: ++49-3641-934755
e-mail: [log in to unmask]
http://dbm.neuro.uni-jena.de
>>> 4) Under Smooth, you must select the kernel size: "Highlight FWHM." We recall that many people double the voxel size. Therefore, if it is 3 as stated above, would we then choose 6 6 6 here?
>> The optimal smoothing will depend on the size of the blobs you expect
>> to find, the accuracy with which data can be aligned across subject,
>> and various other factors. Random field theory is most accurate when
>> the FWHM is a few times greater than the voxel sizes. More smoothing
>> also means you have fewer independent tests, so statistics based on
>> peak height will have fewer multiple comparisons to make (so smaller
>> corrected p values). On the negative side, you can not be so specific
>> about where the differences actually are. I could be wrong, but think
>> most people currently use in the region of 8-12 mm FWHM of smoothing.
>
>As it so happens, there was an interesting paper reviewing and
>investigating this issue in today's issue of HBM: Ball et al.:
>Variability of fMRI-response patterns at different spatial observation
>scales, Hum Brain Mapp 33, 5: 1155�1171, see
>http://onlinelibrary.wiley.com/doi/10.1002/hbm.21274/abstract
>
>Cheers,
>Marko
>
>--
>____________________________________________________
>PD Dr. med. Marko Wilke
> Facharzt f�r Kinder- und Jugendmedizin
> Leiter, Experimentelle P�diatrische Neurobildgebung
> Universit�ts-Kinderklinik
> Abt. III (Neurop�diatrie)
>
>
>Marko Wilke, MD, PhD
> Pediatrician
> Head, Experimental Pediatric Neuroimaging
> University Children's Hospital
> Dept. III (Pediatric Neurology)
>
>
>Hoppe-Seyler-Str. 1
> D - 72076 T�bingen, Germany
> Tel. +49 7071 29-83416
> Fax +49 7071 29-5473
> [log in to unmask]
>
> http://www.medizin.uni-tuebingen.de/kinder/epn
>____________________________________________________
|