 |
 |
Dear Steven, >I have read the archives on Grey Matter thresholds, but I am still confused. > >First, is the Global Mean the average of _all_ the voxels in the matrix or just >the brain voxels ? The general idea is that it should be just the brain voxels. >If the Global Mean is just the brain voxels, then how does >SPM know which voxels to include in the calculation of the Global Mean. > The following paragraph pertains to how intracerebrality is defined for the purpose of calulation of the globals. This is NOT that same way in which intracrebrality is defined when it comes to confining the calulation of statistics. I think some of your confusion may come from the unfourtunate coincidence that the number 0.8 happens to figure both when calculating global flow, and when setting the threshold for confining the calculations. As you correctly say, in order to calculate the "global mean" SPM need some way of figuring out which voxels are within the brain and which are not. Therefore an initial "global" mean is calculated as the mean of all voxels of the image volume. This will as you understand contain also lots of extracerebral voxels (within the bounding box typically used in SPM roughly half of the voxels are extracracerebral) yielding a rather low value. As a next step it finds all voxels which have an intensity value larger than 0.8 of the initial mean, and these voxels are assumed to be intracerebral. Finally the mean of these "intracerebral" voxels is calculated, and this is the measure of "global flow" what will be used in the subsequent analysis. Assuming that the "true" mean activity of the intracerebral voxels is ~50, and that the mean of the extracerebral voxels is ~5 and assuming approximately equal parts of intra and extracerebral voxels within the image volume we see that the inital mean will be ~27.5, and that the threshold for "intracerebrality" when calculating "global activity" will be 0.8*27.5=22. This will obviously include also most of the white matter and some extracerebral voxels. >Let me give an explicit example. Take a PET image with a 128 X 128 matrix and >63 slices. Let the Global Mean be brain voxels only, the Global Mean be 50, and >the grey matter threshold be the default - 80% (0.8). > We are now referring to the threshold for confining the statistical analysis. >Is it then correct that >1) A voxel would have to have a value greater than 40 to be included in the >search volume. > AND a voxel with a value of 25 (50%) of global mean would be excluded from the >search volume > This is correct. Note however that a voxel has to have a value larger than 0.8 times the scan mean in every scan in order to be included. This is effectively a stricter criteria than saying that the average of this voxel should be larger than 0.8 times the average of all scan means. It is roughly equivalent to vM > 0.8*gM+2*vsd where vM is voxel mean, gM is the average global mean and vsd is the voxel standard deviation. >OR is it correct that >2) A voxel with a value of 11 would be included in the search volume because it >is _within_ 80% of the Global Mean. > Nope, this is wrong. >In other words, A) does a voxel need to be greater than 0.8 * Global Mean to be >included in the search volume OR > B) does a voxel need to be greater than 0.2 * Global Mean to be included in the >search volume. > >It appears to me that alternatives 1 and A are what SPM uses, since changing the >Grey Matter threshold to 0.6 leads to more voxels in the search volume. Correct. >If 2 >and B are what SPM uses, then setting the grey matter threshold to 0.6 would >exclude more voxels, since a voxel would have to be greater than 0.4* Global >Mean to be included. > Correct again. >If 1 and A are SPM uses, then 0.8 seems to be a overly stringent threshold since >some sub-cortical areas may only be 0.6 - 0.7 * Global Mean. Not really. Grey matter flow is in the order of 80ml/(min*100ml), whereas white matter flow is in the order of 20ml/(min*100ml). Threshold = 0.8*global mean ~ 0.8*50~40ml/(min*100ml). Now obviously grey matter flow will be underestimated due to partial volume effects, scatter etc but there should still be little risk of discarding actual grey matter regions. If you are worried and really want to see which voxels are included you could redo the analysis and set the F-threshold to p=1.0. Then XYZ.mat will contain all voxels that survive the threshold. >If 2 and B are >what SPM uses, then 0.8 may be overly liberal and include substantial amounts of >white matter. > N.A. Good luck Jesper %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%