Hi all,
I am too quite new to vbm, but I also think it´s essential to smooth (short explanation can be found in K. Friston´s introduction to the Human Brain Function). The results of Christian Gaser script are unsmoothed images.
The question is, how much. The kernel should be greater than image voxel. As I saw most people use kernel between 8-12 mm. Don´t we loose the fine spatial resolution with original voxel sizes around 1 mm3? Another "guide" for kernel size is it should reflect the awaited spatial effect. For fMRI K. Friston shows the spatial scale for haemodynamic response is about 2-5 mm. Are there any similar calculations for vbm? Are there any calculations of how big spatial differences vbm (as implemented using spm approach) can detect? For example in schizophrenia studies I think It is hard to predict the real size of group differences, since post mortem studies give us little data.
But thanks much to Jen for suggestions!
Tomas Kasparek
Dep. of Psychiatry, Masaryk University
Brno
the Czech Republic
______________________________________________________________
> Od: [log in to unmask]
> Komu: [log in to unmask]
> CC:
> Datum: 19.11.2005 00:43
> Předmět: Re: [SPM] Do we need to smooth modulated segments in VBM
>
> Dear Yong Zhang, Joyce Yang & VBM-SPMers -
>
> I am relatively new to VBM analyses as well, so others on the list please
> correct me if the following is bad advice.
>
> I have recently been looking at this issue myself, and have concluded
> that the short answer is YES, you need to smooth if you intend to use
> parametric statistics. There is not a set "rule of thumb" for the
> filter width. How much you smooth should depend on the size of the
> effect you expect to find.
> I recommend reading Jones, D.K. et al., NeuroImage 26 (2005) 546-554.
> Although it is about VBM of DTI data, the same issues apply to VBM of
> structural (SPGR) images.
>
> I used SPMd (available on the SPM2 - Extensions website) to calculate the
> Shapiro-Wilk statistical test for non-normality at each voxel in a 3D-SPGR
> image. Using modulated (non-smoothed) data for a group of 25 normal older
> adults - I found that 74% of the voxels had a significant non-normality
> statistic (alpha=0.05; no smoothing - mGw*; no gray matter threshold
> mask). When the images were smoothed with an 8 mm FWHM kernel, 20% of
> the voxels were significant. When I smoothed with 12mm, 19% of voxels
> were significant.
>
> These numbers improved somewhat when I applied an absolute gray matter
> probability threshold (but still range 11 - 30%) However, do be sure to
> look at your mask images to ensure you have not set the threshold too
> high. In my sample of 65-85 year-olds. (normalized to sample-specific
> template) an absolute threshold of 0.2 dumped most of the voxels in the
> frontal lobe.
>
> You can use the visualization tools in SPMd to see which voxels do not
> conform to the normality assumption required for parametric statistics.
>
> Good Luck,
>
> Jen
>
> Jennifer L. Cox, Ph.D.
> (nee Johnson)
> Postdoctoral Fellow
> Neurological Sciences
> Rush University Medical Center
> Chicago, IL USA
>
> On Nov 13, 2005, at 11:34 AM, Yong Zhang wrote:
>
> > Dear SPMers,
> >
> > I found modulation has a very large and uniform effect.
> > I am not sure if we still need to smooth the modulated
> > segments. If so, what will be a good size for the
> > smoothing kernel, still 12mm or smaller?
> >
> > Thank you!
> >
> > Yong Zhang
> >
> > Radiological Sciences Department
> > St. Jude Children's Research Hospital
> >
>
> >
> > On Nov 14, 2005, at 10:08 PM, Junping Yang wrote:
> >> Dear SPM users and especially VBM users,
> >>
> >> According to the script cg_vbm_optimized.m (provided by http:// >>
> dbm.neuro.uni-jena.de/vbm.html), it seems that there's no >> smoothing
> after the segmentation. And since the templates using in >>
> normalization, which is created by cg_create_templated.m, have >> been
> smoothed by 8mm as default, I'm wondering whether it's needed >> to
> smooth the last image files.
> >>
> >> Two groups with about 30 normal subjects respectively were >>
> involved. I did the analysis following the instruction of VBM >> tools,
> created my own template and use the optimized protocols. At >> last, I
> used the images with and without smoothing respectively in >> statistics,
> and got completely different results. The lateral view >> of four results
> in two studies were attached. It can be easily >> found that the two
> unsmoothed result in two studies are quite >> similar, with few and less
> significant punctate "activation" in >> lateral site and the posterior
> cerebellum of both hemisphere. When >> the images are smoothed, the
> results become much more significant >> (the threshold from uncorrected
> p0.05 to corrected p0.05), and >> show different patterns, (different
> from the unsmoothed results, >> and different between the two studies),
> with more diffused >> "activation" in the brain.
> >>
> >> Do these results mean that the unsmoothed results come from some >>
> kinds of errors introduced in the previous processing, and the >>
> smoothing corrects the errors and thus the smoothed results should >> be
> the right ones?
> >> Or actually the unsmoothed ones are correct and the additive >>
> smoothing just introduces false positive results?
> >> Or the conflicts suggest that there's something wrong in the >>
> previous steps?
> >>
> >> Thank you very much for your intelligence and nice help.
> >>
> >> Best regards,
> >> Joyce
> >>
> >>
> >>
> >>
> >> = = = = = = = = = = = = = = = = = = = =
> >> Joyce J.P. Yang ([log in to unmask])
> >> State Key Lab of Brain and Cognitive Neuroscience
> >> The University of Hong Kong
> >
> >
>
>
>
>
|