As you don't have longitudinal scans for the controls, it will be very
difficult to do a comparison based on the longitudinal data. In
principle, you could just look at the longitudinal scans of the
patients to identify what changed over time, but I would generally
urge a great deal of caution if trying to do this with SPM8. Until
SPM12 is released, I'd suggest maybe trying the ANTS software (
http://www.picsl.upenn.edu/ANTS/ ).
Analysis of longitudinal data is probably best done by within-subject
nonlinear registration of the image series. However, a great deal of
care must be used when doing this, as registration generally involves
aligning one image with another. Swapping the images around will
usually give incompatible results. The image registration community
has known for a long time that asymmetric treatment of data is likely
to lead to artifactual results, although it is only recently - after
some empirical demonstrations - that this has become more widely known
among the rest of the imaging field. For examples, see these papers
apparently written by the Alzheimer's Disease Neuroimaging Initiative:
Paul A. Yushkevich, Brian B. Avants, Sandhitsu R. Das, John Pluta,
Murat Altinay, Caryne Craige, the Alzheimer's Disease Neuroimaging
Initiative
Bias in estimation of hippocampal atrophy using deformation-based
morphometry arises from asymmetric global normalization: An
illustration in ADNI 3 T MRI data
NeuroImage, Volume 50, Issue 2, 1 April 2010, Pages 434-445
Wesley K. Thompson, Dominic Holland, the Alzheimer's Disease
Neuroimaging Initiative
Bias in tensor based morphometry Stat-ROI measures may result in
unrealistic power estimates
NeuroImage, Volume 57, Issue 1, 1 July 2011, Pages 1-4
Note that Dartel works reasonably well for aligning populations of
subject's data together, but it may not be so great for longitudinal
alignment. Some of the reasons for this are:
1) It works at a low resolution, which is not ideal for looking at
sub-voxel changes.
2) It relies on tissue types being segmented, which is probably only
accurate to the nearest voxel. Matching the original images is likely
to be more effective.
3) Longitudinal data can often be very closely aligned, as patterns of
cortical folding etc should not differ much within subject. Therefore
the probabilistic representation of where GM, WM etc is found on
average is not really helpful).
Best regards,
-John
On 10 May 2012 15:03, ethan h <[log in to unmask]> wrote:
>
> Dear experts:
>
> I'm trying to do both cross sectional & longitudinal vbm processing. There are two groups, control (scanned once) and patients (scanned three times).
>
> Although there is longitudinal processing method in vbm8, I cannot figure out how to do comparisons between controls and patients. Can I just use "VBM8: Estimate & Write" to segment and normalize control group images, then compared them with patients images?
>
> Or can I use spm8 DARTEL methods as follows:?
> 1. realign one patients' three images
> 2. realign all three images to the mean image
> 3. use "New Segment" to segment patients' mean images and controls' images
> 4. use results from step3 to create template
> 5. segment patients' all three realigned images, normalize them using flow field parameters created in step 4
> 6. normalize controls' images using flow field parameters created in step 4
> 7. statistical analysis
>
> Thanks!
>
>
>
|
