Hi Donghoon
I think there are two considerations, which sometimes compete. The
first is that a certain amount of spatial smoothness is necessary to
justify the assumptions of random field theory; here I believe the
rule of thumb is at least 3 times the voxel size.
The second is the optimum size smoothing to detect your
activation...for this, the best size smoothing kernel is one that is
the same size as the activation you want to detect. Of course, this
is sometimes tricky to know ahead of time, but if you have a general
idea of the structures it may inform your choice a bit.
As you point out, it is also probably the case that, other things
being equal, a larger smoothing kernel increases the overlap across
subjects (assuming the activations aren't perfectly aligned, which is
probably a safe assumption). In general, unless you are specifically
looking for small structures and not interested in random field theory
for cluster/set level correction, I would tend to stay away from
smaller smoothing kernels...but of course that depends entirely on
what you're looking for. :)
The following page may also be useful:
http://imaging.mrc-cbu.cam.ac.uk/imaging/PrinciplesSmoothing
Hope this helps!
Best regards,
Jonathan
On Wed, Jun 3, 2009 at 11:08 PM, Donghoon Lee <[log in to unmask]> wrote:
> Dear SPMers,
>
> I would like to have some comments about the size of Gausian kernel for
> spatial smoothing.
> I have collected 128x128 EPI images with a 32 channels head coil.
> The original voxel size is 1.72x1.72 x 3.8.
> For spatial normalizing, I resampled data by 2x2x2 voxel size. Then I
> applied Gausian smoothing by 8x8x8 kernel and compared it with 6x6x6 kernel.
> In results, the data smoothed by the bigger kernel show more activation in
> general. However, with 6x6x6 kernel, there are some distinct activation
> spots additionally.
> So, I am suspecting a chance that the smoothing with a larger kernel may
> reduce sensitivity of regional differences. I haven't run a group level
> analysis yet but I am also considering it may affect the group level
> analysis. The bigger kernel may compensate individual anatomical
> difference.
>
> Any comment will be very helpful.
>
> Thanks,
>
> Donghoon
>
> --
> =============================
> Donghoon Lee
>
> Ph.D Candidate
> Department of Psychological and Brain Sciences
> Indiana University
> 1101 East Tenth St.
> Bloomington, IN 47405
> Phone(O): 1-812-856-1776
> E-mail:[log in to unmask]
>
> I prefer keeping in mind even the possibility that existence has its own
> reason for being.
> - Wislawa Szymborska
> =============================
>
>
|