Dear SPMers,
My colleagues and I have recently updated our T1 structural analyses from spm8’s VBM8 to spm12’s DARTEL routines, and are puzzled by the vast differences we are seeing between the two pipelines
in second level grey matter analyses. For example, I ran a simple one-sample t-test to test the for effects of gender with each pipeline, and found widespread effects even at
FWE corrected p<1e-10 with the images generated from the vbm8 pipeline, and only a few small clusters at
uncorrected p<0.5 in the DARTEL pipeline.
Has anyone else seen discrepancies between vbm8 and DARTEL results, or does anybody see anything we are doing horribly wrong in the following outlines of our procedures? So far, the VBM8 results
are more what we’d expect based on existing literature, so we’re more inclined to think that our DARTEL procedure is the culprit.
How we’re running VBM8:
1) VBM8: Estimate & Write using the spm8/toolbox/Seg/TPM.nii, writing out the "Grey Matter" > "Modulated normalized" > "non-linear
only" result.
2) SPM’s smoothing routine with a 12mm kernel
And how we’re running DARTEL:
1)
New Segment, using spm12/tpm/TPM.nii, creating grey & white matter segments for DARTEL import (rc*)
2) "Run DARTEL (existing templates)" using the 6 Template_*_IXI550_MNI152.nii that come with the vbm8 toolbox
(Since our study is quite large and ongoing, and consists of healthy young adults, we normalize to these preexisting templates available with vbm8 rather than continually re-generating templates
from our 1200+ subjects)
3)
DARTEL’s "Normalize to MNI space", leaving the template field blank and selecting the u* flow field from the previous step and the c1* image from segment. Also, we use "Preserve Amount (“modulation”)”, and leave defaults for the rest (smoothing with a 12mm
kernel).
Also, FWIW: as one sanity check, for each subject I calculated the correlations between all non-zero voxels in the brain in the final images from the vbm8 and dartel pipelines, and found them
to be around .95 on average, which seems reasonable.
Any ideas are very much appreciated, thanks! I’m happy to provide more details as needed.
Annchen
~~~~~~~~~~~~
Annchen Knodt, M.S.
Research Analyst
Laboratory of NeuroGenetics
919.684.0277
Duke University