Dear all,
I am trying to compare two groups of subjects using VBM. I computed
modulated normalized images with SPM5 (which gave strikingly good results,
great work John!). Being new to this type of analysis, I perused the list
for a while and came across the issue of global confounds.
If I understand it correctly, the idea in using global confounds in the
model is related to the physical quantity that we are interested in
measuring. A difference in local intensity between two modulated normalized
images can be due to either an original difference in grey matter density,
or to an original difference in the size of the local structure. If we want
to test for effects of local size beyond and above the effects of local grey
matter density, we will have to use a covariate for the effect of grey
matter density as a confound in the model.
And here the confusion begins. In the list, it seems to me that several
approaches have been suggested:
1) Use of the total intracranial volume as a covariate
2) Use of the total grey matter as a covariate
3) Use of the global mean of the grey matter images as a covariate
3) Proportional scaling using one of the above measures.
However, it seems to me that a better approach would be to use the global
mean of the *unnormalized* grey matter images as a regressor in the model.
If you bear with me, I'll try to explain it with a bit of ASCII art.
Let's assume that we have a very simple 2-dimensional brain with just two
rectangular regions of grey matter, and two subjects.
We will examine two possible cases:
1) A real difference in local grey matter volumes between the two subject;
2) A real difference in local grey matter densities with no differences in
local volume.
Here we go:
*******************************************
First case: real difference in local volume
*******************************************
Subj 1, original: tot=100, mean=10
----------------------------------------------
| Region A |
| ---------------- |
| | 10 | 10 | 10 | Region B |
| ---------------- ----------- |
| | 10 | 10 | 10 | | 10 | 10 | |
| ---------------- ----------- |
| | 10 | 10 | |
| ----------- |
| |
----------------------------------------------
Subj 2, original: tot=80, mean=10
----------------------------------------------
| |
| ---------------- |
| | 10 | 10 | 10 | |
| ---------------- ------ |
| | 10 | 10 | 10 | | 10 | |
| ---------------- ------ |
| | 10 | |
| ------ |
| |
----------------------------------------------
Subj 2 modulated normalized: tot=80, mean=8
----------------------------------------------
| |
| ---------------- |
| | 10 | 10 | 10 | |
| ---------------- ----------- |
| | 10 | 10 | 10 | | 5 | 5 | |
| ---------------- ----------- |
| | 5 | 5 | |
| ----------- |
| |
----------------------------------------------
Subj 1 - Subj2.modulated.normalized
----------------------------------------------
| |
| ---------------- |
| | | | | |
| ---------------- ----------- |
| | | | | | 5 | 5 | |
| ---------------- ----------- |
| | 5 | 5 | |
| ----------- |
| |
----------------------------------------------
=> the contrast highlights a (real) difference
in local volume for region B. Same result obtains
if I enter a covariate in the model for the
global mean *before* normalization. Note, however,
that if we use the global mean *after* the modulated
normalization, this will remove our effect of interest.
*****************************************************
Second case: local difference in grey matter density,
with no difference in volume
*****************************************************
Subj 1, original: tot=100, mean=10
----------------------------------------------
| |
| ---------------- |
| | 10 | 10 | 10 | |
| ---------------- ----------- |
| | 10 | 10 | 10 | | 10 | 10 | |
| ---------------- ----------- |
| | 10 | 10 | |
| ----------- |
| |
----------------------------------------------
Subj 2, original: tot=80, mean=8
----------------------------------------------
| |
| ---------------- |
| | 10 | 10 | 10 | |
| ---------------- ----------- |
| | 10 | 10 | 10 | | 5 | 5 | |
| ---------------- ----------- |
| | 5 | 5 | |
| ----------- |
| |
----------------------------------------------
Subj 2, modulated normalized (stays the same):
tot=80, mean=8
----------------------------------------------
| |
| ---------------- |
| | 10 | 10 | 10 | |
| ---------------- ----------- |
| | 10 | 10 | 10 | | 5 | 5 | |
| ---------------- ----------- |
| | 5 | 5 | |
| ----------- |
| |
----------------------------------------------
Subj 1 - Subj2.modulated.normalized
----------------------------------------------
| |
| ---------------- |
| | | | | |
| ---------------- ----------- |
| | | | | | 5 | 5 | |
| ---------------- ----------- |
| | 5 | 5 | |
| ----------- |
| |
----------------------------------------------
=> this difference is NOT due to a real difference
in local volume, but rather to a difference in
local grey matter density. If we enter as a covariate
the global mean, then this is going to eliminate this
spurious effect.
(Tentative) CONCLUSION:
If we want to assess differences in local volume
we should use modulated images AND global means (but
NOT the global grey matter *volume*) as a covariate.
Note the global mean should be computed *before* the
modulated normalization, NOT after.
Note also, (not shown here for brevity), that
performing proportional scaling does not work as well
in these cases as entering a global covariate in the
model: it will produce spurious results, because it
spreads a scaling due to a local effect (region B)
over the whole space, including regions (A) where
no such effect was originally present.
I have also a couple of technical questions:
1) is the total grey matter volume (in mm^3, assuming 1x1x1mm voxels)
computed as the sum of the voxels' intensities of the grey matter image?
2) is the total intracranial volume computed by summing the total grey
matter volume, the total white matter volume, and the total CSF volume?
3) if i didn't output the CFS image in the segmentation process (I used the
default settings), do I need to run the segmentation all over again to
obtain the CSF image and hence the total intracranial volume?
I apologize for the lengthy posting, thanks in advance for any comments and
suggestions.
cheers
giuseppe
|