It should make little difference which parameter set you include as
confounds in the design matrix. For small rotations, each column in
one parameter set can be explained well by a linear combination of the
columns in the other set (plus a constant term).
Best regards,
-John
On 18 June 2013 10:45, Stefan Greulich <[log in to unmask]> wrote:
> Dear all,
>
> thank you John for this thorough explanation. I have only one question left,
> which parameter set to use in as a regressor in the analysis. The 6 movement
> parameters I get after the reorientation or before.
>
> Best,
> Stefan
>
> On 17.06.2013 18:43, John Ashburner wrote:
>>
>> As Helmut pointed out, in the presence of rotations, estimated
>> translations will depend on the origin of the coordinate system. Some
>> people refer to rigid-body registration as linear registration. The
>> behaviour is anything but linear, as a copy/paste of this snippet of
>> MATLAB code will show:
>>
>>
>>
>> % Voxel-to-world mapping of an image
>> M1 = [-2 0 0 92
>> 0 2 0 -128
>> 0 0 2 -74
>> 0 0 0 1];
>>
>>
>> R = spm_matrix(randn(1,6))*0.1; % Some random rigid-body
>> transform
>> %R = spm_matrix([randn(1,3)*0.1 0 0 0]); % Some random translations
>> %R = spm_matrix([0 0 0 randn(1,3)*0.1]); % Some random rotations
>>
>> % Voxel to world mapping after rotation.
>> M2 = R*M1;
>>
>> % This should be the same as R
>> M2/M1
>>
>> % Now we can translate both voxel-to-world mappings by the same amount:
>> T = spm_matrix([40 -10 20]);
>> M1a = T*M1;
>> M2a = T*M2;
>>
>> % And we see that the relative positions of the images have changed:
>> M2a/M1a
>>
>>
>> % Note that the voxel-to-voxel mapping between the images is unchanged
>> M1\M2
>> M1a\M2a
>>
>>
>> Best regards,
>> -John
>>
>>
>>
>> On 17 June 2013 16:21, Stefan Greulich <[log in to unmask]>
>> wrote:
>>>
>>> Dear all,
>>>
>>> here are my movement plots. The voxel size was 3x3x3 mm with 25% gap.
>>> Plot1.png is without reorientation and Plot2.png with reorientation.
>>> The changes are not extreme, but, for example the Y graph changes in the
>>> first run from falling to rising.
>>>
>>> Best,
>>> Stefan
>>>
>>> On 16.06.2013 15:09, H. Nebl wrote:
>>>>
>>>> Dear Christopher and Michael,
>>>>
>>>>
>>>> the voxel size was 3 mm in-plane and slice thickness 3 mm + 15% gap =
>>>> 3.45
>>>> mm. Realignment was onto the mean, if that's of relevance.
>>>>
>>>> Yes, there's not much motion going on indeed. I just took a random data
>>>> set lying around to check what Stefan mentioned, so just for
>>>> visualization
>>>> purpose. I have encountered different RPs depending on the
>>>> (re-)orientation
>>>> of the EPIs previously, but didn't further examine it at that time.
>>>> Basically, if the origin is different, then I'd say the rotation
>>>> parametres
>>>> should be different. Here it's the other way round, same rotation
>>>> parametres
>>>> but different translation parametres. But this might be a logical error,
>>>> depending on which parametres are applied first.
>>>>
>>>> By the way, I've already been wondering what x y z actually stand for in
>>>> the context of realign & unwarp. World space? Voxel space? See
>>>>
>>>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1306&L=spm&F=&S=&P=216973
>>>> But there's no answer so far. Based on the finding above it seems the
>>>> header
>>>> information about the position is not ignored (no idea whether that's
>>>> good
>>>> or bad or irrelevant). Otherwise it shouldn't matter whether volumes are
>>>> translated / rotated before.
>>>>
>>>>
>>>> Best,
>>>>
>>>> Helmut
>>>
>>>
>
|