> We are conducting an analysis of variation of frontal lobe gray matter in
> normals. We are looking to compare male vs. female and left vs. right
> hemisphere, using the VBM optimized protocol. Our population is young (20-
> 40 years) and our images were acquired (3 Tesla) with the following
> sequence: 3D SPGR, TR=9, TE=min, NEX=1, flip angle 20, matrix 256 x 256,
> FOV 24cm, 124 axial slices, slice thickness 1.5mm, in plane
> resolution .9835 x .9835.
>
> The voxel size we are using is 1 x 1 x 1. Our question regards the
> appropriate sized smoothing kernel. It seems that 8mm for creating a
> study- specific template and 12mm for the final images to be analyzed is
> standard. However, given that we are looking for small anatomical
> variation, are smaller kernels ok to use, still leaving valid statistics?
Smaller kernels should produce valid statistics providing that you are
comparing groups rather than single subjects versus groups. Smaller
kernels may influence the interpretation of the results though. Spatial
normalisation is only accurate to within about 10mm in the cortex, so
less smoothing is more likely to show differences that are due to systematic
registration errors rather than true volumetric differences.
You also need to consider how VBM works. It basically only shows differences
at edges, even though any true differences are not necessarily found right
at the edge of the grey matter. Smoothing has the effect of spreading these
differences out so that superimpose in very roughly the same
place as the volumetric difference causing the displacement. It is very hard
to say exactly where the volumetric differences lie, so they can not be
really accurately localised with VBM (or any other method based only on T1
weighted images).
> It was said awhile ago that the smoothing kernel need only be twice the
> voxel size, so given our small voxel size, is a small kernel appropriate?
There are other considerations with multi-subject analyses - i.e. inter-subject
registration error prevents you from really accurately localising any
differences.
>
> We would very much appreciate experts' opinions about how small our kernel
> can be so it doesn't wash out differences we are looking for but still
> allows for valid analyses. Would 4mm for the template and 6mm for the
> final images be appropriate?
I would still suggest using 8mm for the template, as a smoothing of 8mm is
hard-coded in the spatial normalisation software. 6mm for the final images
may be OK for some regions where there is less anatomical variability and
therefore more accurate spatial normalisation, but is not enough for
the cortex.
Best regards,
-John
--
Dr John Ashburner.
Functional Imaging Lab., 12 Queen Square, London WC1N 3BG, UK.
tel: +44 (0)20 78337491 or +44 (0)20 78373611 x4381
fax: +44 (0)20 78131420 http://www.fil.ion.ucl.ac.uk/~john
|