I generally regard the "average neighbors" smoothing as a legacy
method, to be used only if you need to compare to a previous study
that also used it.  The geodesic gaussian smoothing method in caret5
uses a gaussian kernel based on geodesic distance directly, so you
don't need to futz with the parameters in order to get the same
spatial smoothing on a different density mesh.  Furthermore, workbench
improves upon that slightly, for surfaces that have highly irregular
node areas (like freesurfer native meshes), with the geodesic gaussian
area method, which additionally takes node areas into account when
calculating the weights.  If you want a specific FWHM or gaussian
kernel, I recommend these two methods.  Due to its shoehorning into
caret5's existing metric smoothing command, geodesic gaussian
smoothing ignores the "strength" parameter, and I recommend using a
single iteration, specifying the desired sigma of the gaussian kernel
with  the "-geo-gauss" option.  Beware, the "GAUSS" method is
something else entirely, and the "-gauss" option is similarly
unrelated.

Unfortunately, the average neighbors method's effective FWHM or
gaussian kernel can't be estimated only by the iterations and
strength, it also depends on the mesh density (basically, on how long
each edge is, on average).  I don't know of a robust formula to
estimate the effective kernel, it would need an estimate of the
effective gaussian kernel for one iteration, which varies by the
strength parameter.  There should be a gaussian kernel that generates
very similar smoothing as long as you do more than 1 or 2 iterations
with average neighbors, since repeated convolution of any smoothing
kernel approximates a gaussian kernel.  It really boils down to: if
you want to know the effective gaussian kernel, as far as I know, you
need to test empirically against geodesic gaussian and compare, or use
a metric with a single nonzero value and use the result to measure the
size of the kernel.

As for translating an existing average neighbors setting to a
different density mesh, multiplying the iterations by the ratio of
nodes should, in theory, work out to a similar sized effective
gaussian kernel.  This is because the node spacing varies with
approximately the square root of the number of surface nodes (given
the same total surface area), and the effective gaussian kernel varies
with the average node spacing times the square root of the number of
iterations.

Smoothing immediately before TFCE is done largely because TFCE is very
sensitive to smoothness (since it uses contiguous above-threshold
areas as part of the calculation, and noise has a very big impact on
contiguity).  It may be equivalent to smooth the data before
generating the t-map, but the only place that the smoothness really
matters, as far as I know, is in TFCE.

Tim

On Fri, Jul 20, 2012 at 2:02 PM, Donna Dierker <[email protected]> wrote:
> Joern Diedrichsen played around with this years ago and found four average
> neighbors iterations to work well with caret-style surfaces:
>
> www.icn.ucl.ac.uk/motorcontrol/download/Caret_surface_statistics.pdf
>
> But he probably smoothed the mapped fMRI before doing any sort of t- or
> f-test.  We typically are working with depth, which is already pretty
> smooth, but decided to do this small amount of smoothing on the resulting
> t-/f-maps (real and randomized).
>
> We recently started doing this stuff on the 164k mesh, and we use 9
> iterations for that purpose (4*164k/74k).  (Tim Coalson said some
> nonlinearities more or less cancel each other out, so it's mostly
> proportional with the number of vertices.)
>
> Read what Smith & Nichols say about smoothing in the TFCE paper and form
> your own conclusions, based on your data.  The most important thing is to
> be consistent (smooth all subjects/sessions the same amount/way).
>
>
>> Hi Donna,
>>
>> I have used the TFCE method to generate some statistical meaningful
> results on cortical asymmetries. Thanks a lot for your help.
>> Currently, I have a question regarding to step of smoothing t-map with 4
> iterations at 0.5 strength using average neighbors method. I found that
> this step indeed improved the statistical results. What''s
>> difference between smoothing t-map and original surface attributes such
> as sulcal depth or surface area? Why we need smooth t-map instead of
> surface attributes? And also, how the with 4 iterations at 0.5 strength
> using average neighbors method relate to the Gaussian kernel size or
> FWHM on my surface mesh with 160K vertices? Thanks.
>>
>> Regards,
>>
>> Gang
>> _______________________________________________
>> caret-users mailing list
>> [email protected]
>> http://brainvis.wustl.edu/mailman/listinfo/caret-users
>>
>
>
>
> _______________________________________________
> caret-users mailing list
> [email protected]
> http://brainvis.wustl.edu/mailman/listinfo/caret-users
_______________________________________________
caret-users mailing list
[email protected]
http://brainvis.wustl.edu/mailman/listinfo/caret-users

Reply via email to