I'm not John but have an opinion on this. If you modulate for every warp, results will be in units of volume. I find "volume"/mm3/cm3 easier to interpret and easier to explain to non-imaging journal reviewers. If I can say "Group X had an average gray matter volume change of +8 cm3 in this region over two years while Group Y had an average volume change of -3 cm3" that is quite tangible.
On a related note, I did some cross-sectional analyses with DARTEL and compared them with published results we got from the same data set with SPM5. I found that regions were smaller and less significant with DARTEL. Blobs did appear in the same locations and with nearly identical peak voxels as with SPM5, which was reassuring... but some were sub-threshold when previously they were reportable (and reported!). Smaller cortical significant blobs with DARTEL do make sense to me, since the individual brains appear squashed into more similar space. I was hoping there would be compensation by higher T or smoothness would be "rougher" in those regions, making the smoothness-corrected cluster-level significance the same as it was with SPM5... but it wasn't. Clusters were simply less significant with DARTEL. This makes me worry about Type 2 error with DARTEL (or Type 1 with SPM5?), but it makes me feel pretty good about my positive results recently obtained from longitudinal DARTEL. In addition, statistical differences in my longitudinal DARTEL analyses were strictly in the direction of our hypotheses :) often with thousands of voxels and a handful of significant clusters in the hypothesized direction, and not a single supra-threshold voxel in the contrast in the other direction.
-Dana
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Neil Chatterjee
Sent: Friday, May 22, 2009 11:27 AM
To: [log in to unmask]
Subject: Re: [SPM] R: [SPM] Longitudinal DARTEL spm8
Hi John,
(hopefully) quick question about the method described below: why
modulate again in step f? It seems like deformations from the
intra-subject warping is the interesting part and inter-subject warping
is just to put things in a common space for logistical ease of doing
stats. So why input inter-subject warping info into the images by
modulating again?
Regards,
Neil
Neil Chatterjee
Stanford Systems Neuroscience and Pain Lab
[log in to unmask]
John Ashburner wrote:
> The suggestion I made was about how to use DARTEL to do the intra-subject
> alignment, so I was basically just answering the question. I have no
> empirical evidence either way, but (for the reasons I mentioned) I suspect
> that HDW may do a marginally better job with most serial scans.
>
> Best regards,
> -John
>
> On Friday 22 May 2009 11:15, Benetti, Stefania wrote:
>
>> Thank you John for your answers and suggestions. However, I am not sure to
>> understand what you are suggesting about HDW approach that Kipps used.
>>
>> "In general, I think I would still suggest the HDW approach that Kipps and
>> others have used...For inter-subject alignment, the residuals are rarely
>> i.i.d. Gaussian, which is why I chose to align tissue class data
>> instead.... However, the segmentation errors may be relatively large
>> compared with the volumetric differences between the serial scans, which
>> would make the DARTEL approach less accurate"
>>
>> My understanding was that the procedure you suggested to Reinders
>> (https://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0804&L=SPM&P=R48484)
>> consisted in an adaptation of Kipps' approach for DARTEL, in which the
>> within subject alignment approach replaced the HDW. Am I completely wrong?
>> Did you mean that this longitudinal DARTEL procedure may be less accurate
>> than the procedure that Kipps actually used?
>>
>> Many thanks again
>> Stefania
>>
>>
>> -----------------------------
>>
>> Stefania Benetti
>> King's College - Institute of Psychiatry
>> Neuroimaging Section
>>
>>
>>
>>
>> -----Messaggio originale-----
>> Da: John Ashburner [mailto:[log in to unmask]]
>> Inviato: gio 21/05/2009 13.50
>> A: Benetti, Stefania; [log in to unmask]
>> Oggetto: Re: [SPM] Longitudinal DARTEL spm8
>>
>>
>>> We pre-processed a small longitudinal dataset (26 subjects,T1=baseline
>>> T2=follow-up) using DARTEL and the procedure suggested in:
>>>
>>> https://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0804&L=SPM&P=R48484
>>>
>>> Briefly,
>>> a) both T1 and T2 manually reoriented. T2 co-registered to T1,no
>>> re-slicing. b) segmentation of both T1 and T2 images (section 1.1 DARTEL
>>> manual). c) create within-subject template (smoothing parameter to NONE)
>>> d) generate modulated warped GM an WM using the within-subject flow
>>> e) create an inter-subject template
>>> f) generated modulated warped(mwmwc1T1 and mwmwc1T2) GM and WM using the
>>> inter-subject flow fields. g) smoothing and statistical analysis using
>>> both a flexible factorial design.
>>>
>>> However when we pre-processed the same database using optimised VBM for
>>> serial scans and then the same statistical analyses, we obtained a
>>> completely different result. GM changes were found in regions where no
>>> significant effects were detected with the DARTEL approach.
>>>
>>> Would you expect to find such a difference?
>>>
>> I would expect different models to give different results, so I'm not
>> really surprised. I would suggest checking out the contrast images
>> generated from the GLM to see if the general trends are similar.
>>
>>
>>> Could this difference in terms
>>> of localisation be attributable to differences in normalisation?
>>>
>> Very likely, and also differences in tissue classification.
>>
>>
>>> Is it
>>> sensible to rely on the DARTEL approach rather than the optimised one?
>>>
>> I haven't tested the various approaches to know what works "best", and the
>> most sucessful approach is likely to be dependent on things like the
>> contrast in the images, the image artifacts and the stability of the
>> scanner. However, I would expect that some form of within subject alignment
>> approach may provide more sensitivity to differences. For longitudinal
>> analyses, you are identifying tiny volumetric differences of the order of a
>> percentage or so, so the details really do matter.
>>
>> In general, I think I would still suggest the HDW approach that Kipps and
>> others have used. There are issues with HDW, which relate to the algorithm
>> fully converging and it is also asymmetric (so registering early with late
>> will give different results from doing it the other way around), but it is
>> probably the more accurate of the SPM procedures to use for longitudinal
>> studies. A histogram of the difference between the registered images should
>> approximately indicate that the residuals are i.i.d. Gaussian, which would
>> make the mean-squares difference (used by HDW) a suitable objective
>> function to use. For inter-subject alignment, the residuals are rarely
>> i.i.d. Gaussian, which is why I chose to align tissue class data instead.
>> However, the segmentation errors may be relatively large compared with the
>> volumetric differences between the serial scans, which would make the
>> DARTEL approach less accurate.
>>
>>
>>> One more question about step d) in the procedure mentioned above. Is this
>>> step necessary? I am probably missing something, but I was wondering if
>>> warping the T1 and T2 images using the within-subject flow fields, and
>>> then warping the obtained images again with the inter-subject flow fields
>>> (step f) may reduce local differences we are actually interested in
>>> since we are dealing with serial scans.
>>>
>> The transforms could be composed and then used, but I'm not sure how much
>> difference it would make in practice.
>>
>> Best regards,
>> -John
>>
|
