Dear Christian,
thanks a lot again for your answer.
I decided to omit the residuals-images approach. I had originally chosen it because I feared that differences between TIV-estimation of the different scans would affect my statistics, but it turned out to explain only very little variance.
More importantly though, I do find negative values in the segmented grey-matter-images (as well as in the segmented white-matter images). This problem did not arise from the use of the residual images.
I attached a pdf-file to this email with screenshots of the images in fslview and their minimum and maximum intensity values (fslstats).
It is only a small number of voxels that occur at the "edge" of the GM.
These negative vales only occur when I use the longitudinal segmentation, not when I segment all images separately (see PDF)! So it must arise from one step in the longitudinal segmentation.
I noticed that the average images already have negative values (see pdf) and assumed that this was behind the negative values in my data.
But I do not understand how these arrive from averaging two images without negative values!?
In order to get a better overview of how these values are distributed, I thresholded the images to min=0 and and max=0. Screenshots of the respective outputs can be found in the pdf-file attached to this email as well.
I looked at the data of 3 colleagues (MPRAGE as well as MP2RAGE) preprocessed with cat12 -longitudinally and the same thing occurs in their data sets.
Maybe we have a false setting in the batch? (i will attach my batch to this email as well)
I would be very happy about your input on this.
Thank you so much!
Marie.
----- Original Message -----
From: "Christian Gaser" <[log in to unmask]>
To: "Marie Uhlig" <[log in to unmask]>
Cc: [log in to unmask]
Sent: Friday, October 21, 2016 10:33:49 PM
Subject: Re: delta grey matter images
Dear Marie,
sorry for the delay, I was out of office in September and forgot to
answer…
On 27 Sep 2016, at 1:27, Marie Uhlig wrote:
> Dear Christian,
>
> Thanks a lot for your reply and explanation.
>
> The reason why I did not use the flexible factorial design, was that I
> understood this to model the variance for every subject separately and
> therefore making any
>
> covariates obsolete, since individual variance would already be
> accounted for.
>
> Maybe I missunderstood this one!?
>
> I also have two follow-up questions:
>
> 1) What is the interpretation of the negative values in the grey
> matter images? I assume they arise from the registration to the
> average image, but the process is not
There should be never occurring negative values in your segmentations. I
assume that this is caused by your use of the residual images. I still
don’t understand the way and the motivation for your residual images.
Is there any reference for that? Anyway, I wouldn’t make it too
complicated and would either try the flexible factorial with the single
GM segmentations for each time point or a full factorial model for the
difference images. The latter might be easier if you also have
covariates. However, you have to create a GM mask for limiting your
analysis on GM because you cannot use an absolute threshold (see my last
mail).
Best,
Christian
>
> quite clear to me yet.
>
> 2) Do you think the negative values in the data provide a problem for
> creating the difference image, in the sense that when two negative
> values are substracted from
>
> one another, this results in an increase rather than a decrease ?
> (i.e. (-3)- (-2)=(-3)+2=1) Or would that maybe even be desirable? I
> guess this also depends on the
>
> above answer.
>
> Thanks a lot for your help!
>
> Best Wishes,
>
> Marie.
>
> ----- Ursprüngliche Mail -----
> Von: "Christian Gaser" <[log in to unmask]>
> An: [log in to unmask], "Marie Uhlig" <[log in to unmask]>
> CC: "Christian Gaser" <[log in to unmask]>
> Gesendet: Montag, 26. September 2016 23:44:32
> Betreff: Re: delta grey matter images
>
> Dear Marie,
>
> the motivation why you are using residual images to consider TIV is
> not clear for me. If you are interested in an analysis of an
> interaction between time and group I would suggest to use one of these
> approaches:
> 1. Use a flexible factorial model with the factors subject, group, and
> time and model an interaction between factors 2 and 3 (group and time)
> and a main factor 1 (subject). Use TIV as nuisance variable. There is
> an example in the CAT12 manual at page 24.
> 2. Calculate difference images between the two time points and compare
> the difference image using a two sample t test. Use the mean TIV of
> both time points as nuisance parameter. Please don't use an absolute
> threshold for masking because the difference images also contain
> meaningful information < 0. I would suggest to create a mask based on
> the thresholded mean image of GM of one time point.
>
> Both models use modulated images and TIV as nuisance parameter and
> allow to test for different time curves between both groups
> (interaction time x group).
>
> Best,
>
> Christian
>
> On Sun, 25 Sep 2016 03:33:28 +0200, Marie Uhlig <[log in to unmask]>
> wrote:
>
>> Dear SPM-Mailing List,
>>
>> I am running a VBM-Analysis investigating the effect of an
>> intervention on two different groups.
>>
>> I have used the longitudinal segmentation of the cat12-toolbox and i
>> am working with the (s)mwp1-(grey matter)-images.
>>
>> I used a multiple regression with the TIV covariate (separately for
>> each timepoint) in order to obtain the residual images (Res_*.nii),
>> that have been corrected for any effect of modulation.
>>
>> In order to model the difference of these images i have calculated
>> the delta (timepoint2-timepoint1).
>>
>> However, i am running into the problem, that negative values are
>> present in my data before subtraction already. Therefore a decrease
>> in GM may result in larger values in the delta-image instead of
>> smaller
>>
>> (eg. (-3) - (-2)= (-1) ).
>>
>>
>> Does anyone have any idea on how to circumvent this?
>>
>>
>> I assume using working with the unmodulated, segmented grey matter
>> images would be the optimal choice here, but are there other
>> approaches when using the modulated data?
>>
>>
>>
>> Thanks a lot for your help in advance!
>>
>> All the best!
>>
>> Marie.
|