Dear list,
I was wondering whether this discrepancy has been noted before and discussed elsewhere (a brief search didn't yield any more results at least).
After noticing the difference in smoothed images, I'd like to follow up on Robert's email with the following observation:
When I use a call to spm_smoothkern (incl. the last SPM8 update) to print the kernel weights to the command line, like this
spm_smoothkern(2, -6:6, 1)'
I get the following weights:
0.0000
0.0000
0.0000
0.0026
0.0447
0.2417
0.4221
0.2417
0.0447
0.0026
0.0000
0.0000
0.0000
Now, my understanding of the FWHM concept is that at full width (that is to say the distance between the two points with equal distance from the max and equal value), the value should be half the maximum. As is clear, 0.2417 is (significantly) larger than half of 0.4221 (= 0.2111). According to what I can tell, this is more like a kernel of FWHM 2.2.
In essence, I believe that the current implementation "over-smoothes" images. To further see to what extent this occurs, I computed the ratios between the values at FWHM point to maximum for different kernels using the following code:
kernels = 1:0.2:4;
kernel_ratios = zeros(size(kernels));
for kc = 1:numel(kernels)
k = spm_smoothkern(kernels(kc), [-kernels(kc)/2, 0], 1);
kernel_ratios(kc) = k(1) / k(2);
end
When plotting the results with
plot(kernels, kernel_ratios)
I can observe that the smaller the desired kernel's FWHM, the larger the discrepancy.
Now the question is: Is this *desired* or is this some kind of flaw in the new code of SPM8??
Looking for answers...
/jochen
|