Dear Jaap,
> We are currenlty planning to make a voxel-by-voxel statistical comparison
> of ADC maps between healthy subjects and patients with epilepsy.
> The approach we use is similar to that described in:
>
> Diffusion tensor imaging of cryptogenic and acquired partial epilepsies.
> Rugg-Gunn FJ, Eriksson SH, Symms MR, Barker GJ, Duncan JS.
> Brain. 2001 Mar;124(Pt 3):627-36.
> http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?
> cmd=Retrieve&db=pubmed&dopt=Abstract&list_uids=11222461
>
> Using the mean ADC-map and the standard deviation ADC-map of the healthy
> volunteers we are able to find voxels in the ADC map of each patient
> individually that significantly (P < 0.001) deviate from the healthy
group.
> However, we also want to correct for multiple comparisons, using a
> corrected P value P < 0.05). (with reference (Friston et al 1994))
We are looking at this problem in our lab, also using ADC maps. However, we
look at the whole brain (we have no anatomical ROI). We found that (as also
found by others, like Nichols & Hayasaka, Statistical Methods in Medical
Research, 12(5): 419-446, 2003) random field theory corrections tend to be
rather conservative unless you have a large number of subjects in your
reference sample (ours has 40).
Our approach to the problem was that of calibrating the degrees of freedom
of the random field using resampling methods (in the resampling literature
there is some experience on calibration, especially in the bootstrap field).
Besides giving better power, this approach also has the advantage that it
also gives you information on the extent to which types of images that
aren't often tested with SPM are actually adequately modelled by random
fields. It turned out that in our ADC images, if you look for signal
hypointensities the random field tests are too conservative (that is, ADC
maps have shorter tails than in the random field model). Signal
hyperintensities are modelled about right. However, this appears to depend
on the way you segment your images, as segmentation influences the extent to
which partial volume effects change the signal distribution (an additional
reason in my view to use calibration). It also depends on the quality of
your estimate of the variance from the reference sample (the results just
cited are valid for smoothed variance estimates).
Our work has been submitted last December, we haven't (alas) heard anything
yet. In case, you are welcome to contact me separately on this.
>
> Our initial thought was to use spm_est_smoothness.m & spm_resels.m
> Using a temporal lobe mask, we should be able to obtain a FWHM estimate
for
> the temporal lobe area of each ADC-map, and after that a resel count for
> the temporal lobe. Then, we would apply a Bonferroni correction with the
> number of resels in stead of the number of voxels.
I am not sure I remember well, but I think there is actually a warning
somewhere that doing a Bonferroni correction with the resels count does not
give the same result as the random field theory test. It would be
interesting to have some simulations clarifying what happens when you do
this.
Roberto Viviani, PhD
Dept. of Psychiatry
University of Ulm
Germany
|