>Hi: I was using SPM2 to compare the PET images of a family of patients
against normal subjects but am having trouble with the results. Wonder if
anyone may have some input and wisdom on what I have done wrong. Thank you
much in advance. Here is what I have done thus far:
>1. Coregister each patient's/normal PET with his/her MRI
Yes - but best to use "coregister only" - saves you one interpolation step. In SPM2 this transformation goes into the PET.mat file and gets called up every time you deal with the image.
2. Spatially normalize patient's MRI with the normal template in SPM2
Why not. The SEGMENT module in SPM5 would do a better job (Ashburner and Friston 2005, Unified Segmentation). If you have got lots of subjects making your own template (in analogy to VBM) out of lots of controls and patients may be better, particularly if your patients are anatomically different - which seems to be the case. There is a lot of information on "individualised" templates in the archives.
3. Write the transformation back to the PET images.
I guess you mean that you apply the spatial transformation parameters to the PETs - that's correct.
4. Express the values of PET SUV value as a ratio of basal ganglia (I
used PMOD to do this).
I can't comment as I don't know which tracer you use. If FDG, the ratio to basal ganglia may not be optimal (e.g. Yakushev I et al. Psych Res 2008, Yakushev I et al. Neuroimage 2008, Borghammer P Neuroimage in press). If you don't use internal ratios, using SUVs may also be unnecessary if you use ANCOVAs.
5. Using SPM2 --> PET --> Two sample T test
6. Compare a single patient against 9 normal subjects
Not a lot... any chance of increasing the number of normals?
7. Setting: No global normalization, no mean scaling, threshold absolute at
0.4 (and I tried from 0.3 to 0.8).
It may be preferable to skip 4. and go for ANCOVA (per condition) here. You place a lot of trust in your scaling and then treat the result as completely trustworthy absolute values - there are so many variables that that's rarely a prudent thing to do. I think what we arrived at (for ligand PET) should be described in Hammers et al. Brain 2002, 2003 - including a method to combine low relative thresholds with masks constraining the maximum extent in which SPM goes testing. I wouldn't generally recommend absolute thresholds.
9. Define t-contrast [-1 1 0] as I want to see which parts of the patient's
brain has lower uptake than normal.
Not sure where the third column comes from...
10. Set FDR p<0.05, Thresholds I tried from 0 to 200.
Strictly speaking if you try multiple (extent?) thresholds you should adjust for that. I've never seen the rationale for extent thresholds - if you're looking for a change in a specific area the resulting cluster may be very small indeed - we had very small ones in temporal lobe epilepsy bang in the hippocampus, not in controls, and in some patients confirmed in terms of relevance with other techniques... It all depends on your question.
11. Save the overlay image --> Overlay with MRI.
Hm - the resulting SPM? Is this a single patient, or a group?
>As the attached jpeg image shows, the result is strange that it does not
follow any anatomic structure, but rather outlines the brain. I expect
difference between the patient and the normal in large areas (e.g., the
entire temporal lobe or more). I am quite puzzled... particularly I might
have missed something quite fundamental. Anyone has any idea?
>Just FYI: some of my patients have cortical atrophy and venticular
dilatation. Would this affect the calculation?
Very, very likely.
Hope this gives some starting points!
All the best,