It's all to do with the masking that is applied in the first level
analysis. The grey regions you see in the normalised images actually
have a value of zero, whereas the black parts have a value of NaN.
The segmentation uses an older approach to spatially normalise images,
which just involves resampling the data. Values outside the FOV are
set to NaN because they are undefined.
The Normalise to MNI approach is done slightly differently. It pushes
the voxels to their new locations and keeps a count of how many voxels
arrive in each voxel of the spatially normalised images. The pushed
voxel image and the count image are then smoothed, and the former is
divided by the latter. Actually, to try to reduce people's concerns
about the swirly stuff where there is missing data, the division is by
the voxel count + 0.001. This causes missing data to tend towards
zero, without biasing the other parts too much. NaNs are ignored in
the pushing, so the background regions end up having values close to
zero.
The general idea behind the newer procedure is that it should try not
to lose signal. For example, if a simple downsampling approach is
used to transform an image of 1 mm resolution, down to one of 2mm
resolution (by sampling voxels every 2 mm), then it is possible to
lose up to 87% of the original signal.
Best regards,
-John
On 6 June 2013 09:20, Lawson, Rebecca <[log in to unmask]> wrote:
> Thanks for your replyJohn, i really appreciate it.
>
> Just to be clear. These aren't new subjects that I? Resting differently and adding in, these are the same subjects but just a new analysis. So i'm normalising the first level contrast images in exactly the same way that I normalised the functional (and structural) images for these same subjects in a previous analysis. I only see this inflation of the volume when normalising the contrast images though. Also, the normalised cons don't look 'flowery', there is just a continuation (beyond the brain in some/all panes) of uniformly grey voxels (basically to the edge of the FOV) which is inflating the mask at the second level.
>
> I can go back and run the first level on normalised images as you suggest. I guess my curiosity just got the better of me and I wondered why normalising in exactly the same way did not cause these effects for functional and structural images, only the contrast images.
>
> Cheers,
> B.
>
> _______________________________________
> From: John Ashburner [[log in to unmask]]
> Sent: 05 June 2013 22:36
> To: Lawson, Rebecca
> Cc: [log in to unmask]
> Subject: Re: [SPM] Dartel normalise contrast mask
>
> In general, it is safest to use the same (or similar) preprocessing
> for all your subjects. Whether you do the GLM after or before
> spatially normalising should not really make a large difference.
> There will be a small discrepancy due to the REML estimation of the
> temporal correlations, but the thing causing the main difference will
> be the behaviour of the masking.
>
> The easiest thing would be to use the "Run Dartel (existing template)"
> option (using the previously computed Template data from the other
> subjects) on the additional subjects, followed by the Normalise to MNI
> space option, and then the first level analysis to get the contrast
> images.
>
> Generally, the flowery/swirly artifacts from normalising to MNI space
> will occur for regions of the head where no data was available in the
> original scans.
>
> Best regards,
> -John
>
>
> On 5 June 2013 19:00, Lawson, Rebecca <[log in to unmask]> wrote:
>> Hello,
>>
>>
>>
>> I’ve used Dartel to normalise some functional images and then taken these to
>> first level analysis, then smoothed the contrast images and taken them to
>> second level analysis. So long as I set the following during smoothing :
>> matlabbatch{1}.spm.spatial.smooth.im = 1;%set mask to 1 to avoid crazy mask
>> my second level mask looks normal.
>>
>>
>>
>> I now have a new first level analysis conducted in native space. I am
>> attempting to normalise and then smooth the 1st level contrast images using
>> Dartel. I have tried both the ‘normalise2mni’ routine and also the ‘combine
>> deformations’ route (followed by a separate smoothing step) to achieve the
>> normalisation of these contrast images. In both cases the normalised
>> contrasts (and hence the mask created at the second level) are massive! The
>> voxels around the brain (that are NaN in the functional images) are all
>> given numbers very close to zero, and hence appear grey right to the edge of
>> the FOV. Using the ‘normalise2mni’ routine this inflation occurs in x, y and
>> z planes. Using the ‘combine deformations’ route this inflation only occurs
>> in the x plane (left to right).
>>
>>
>>
>> I have already searched the spm archive and this “greying” of the contrast
>> image beyond the brain looks different to the “flowery” effects that other
>> people have reported after normalising contrast images using Dartel.
>>
>>
>>
>> Any ideas folks?
>>
>>
>>
>> B.
>>
>>
>>
>> -------------------------------------------------------
>>
>> Dr Rebecca Lawson
>>
>> UCL Institute of Cognitive Neuroscience
>>
>> 17 Queen Square
>>
>> London
>>
>> WC1N 3AR
>>
>>
>
>
|