Overlaying 3dMSS results on T1 - large pixels

[Annen]

Hello AFNI experts,

I am doing some analyses with 3dMSS, on data preprocessed with FSL. Now I would like to visualize the results on the anatomical images, the group average T1 has large pixels, dispite the voxel size of 1x1x1. I would like a smoother image, as the original image on the right:

Is there anything I can do to improve the visualisation?

Thanks for your help,
Jitka

[ptaylor]

Hi, Jitka-

Just to be sure, the issue is that the voxels themselves are visible here? (I'm curious what happens if the axial image in the viewer on the right is increased to the same size---are the voxels not visible there?)

And are those voxels really 1mm isotropic in the lefthand image? What is the output of:

3dinfo -ad3 DSET

... on that dset?

--pt

[Annen]

@ptaylor

3dinfo -ad3

Hi Paul,

yes, I am sure the voxels are 1mm isotropic. This is the output of 3dinfo:

3dinfo -ad3 Avg_xxx_correctedAff_Vox.nii 
** AFNI converts NIFTI_datatype=4 (INT16) in file Avg_xxx_correctedAff_Vox.nii to FLOAT32
     Warnings of this type will be muted for this session.
     Set AFNI_NIFTI_TYPE_WARN to YES to see them all, NO to see none.
1.000000	1.000000	1.000000

There really seems to be an issue with the visualisation in the afni viewer, and I don't think the voxels visible right now corresponds to the voxel size. Here is a screenshot of the same image and approximately the same size in both the afni viewer and MRIcron.

Best,
Jitka

[ptaylor]

Hi, Jitka-

I would be pretty surprised if there were a a visualization issue with the GUI. By default, it shows things at actual resolution; the data are discrete voxels. The image on the right (non-AFNI GUI) looks quite blurry to me, in a way that the actual representation of the data is affected, and I don't know why I can't see the voxels.

By default, the AFNI GUI does not blur/interpolate/smooth the data. If you want to blur things, you do the following in the GUI:

• click "Define Datamode" (the button just a bit above where the "Underlay" menu is)
• at the top of the new menu, click "Warp Ulay on Demand"
• in that top section, choose whatever Ulay Resam mode you want for resampling/blurring the data
• choose a given resampling size.

In doing so, the data can be made to look smoother.

I'd be curious to take a look at this dataset, if it can be shared. I will message you about that.

--pt

[ptaylor]

Thanks for sharing the dataset. There are a few things happening here.

I'm going to compare the shared dset, referred to and renamed to be shared_t1.nii.gz, that is visualized above, with the distributed template, MNI152_2009_template.nii.gz.

Let me first share picture of them for a similar, middle-ish slice of anatomy, where MNI is on the left and "shared" is on the right (sidenote: it is worth noting for a point below that it appears that the crosshairs are locked to the same coordinate here, but in fact they are not; these two panels are unlocked and I have just placed the crosshairs in a similar-ish anatomical location for each, to match visually):

This is just raw visualization in the GUI. The one on the left is sharper and has more tissue contrast, fine, but the voxel discreteness on the right is more apparent. That latter question is what got us started here.

First, if we compare voxels sizes, both, we see they are equal (1mm isotropic):

3dinfo -ad3 -prefix MNI152_2009_template.nii.gzshared_t1.nii.gz  
1.000000	1.000000	1.000000	MNI152_2009_template.nii.gz
1.000000	1.000000	1.000000	           shared_t1.nii.gz

But, if we look at matrix dimension size, we see a big difference:

3dinfo -n4 -prefix  MNI152_2009_template.nii.gz shared_t1.nii.gz 
193	229	193	1	MNI152_2009_template.nii.gz
91	109	91	1	           shared_t1.nii.gz

So, even though the brains appear to be the same size when we have the axial window edges lined up, in fact we are quite zoomed in when looking at the shared dataset.

We can compare the autobox size and output the min/max extent of the brain along each axis, and we see that for MNI:

3dAutobox -extent MNI152_2009_template.nii.gz 
++ 3dAutobox: AFNI version=AFNI_24.2.02 (Aug 16 2024) [64-bit]
++ Auto bbox: x=25..168  y=25..205  z=6..160
Extent auto bbox: R=-72.000000 L=71.000000  A=-73.000000 P=107.000000  I=-72.000000 S=82.000000

... the RL brain extent is about 143mm, AP extent is 180mm and IS extent is 154mm.

The autobox extent for the shared dset is given by:

3dAutobox -extent shared_t1.nii.gz
++ 3dAutobox: AFNI version=AFNI_24.2.02 (Aug 16 2024) [64-bit]
++ Auto bbox: x=10..79  y=10..98  z=1..75
Extent auto bbox: R=-80.000000 L=-11.000000  A=-99.000000 P=-11.000000  I=2.000000 S=76.000000

... so the RL brain extent here is only (PT: edited to do the algebra of extents-math correctly now) 69mm, the AP extent is 88mm and IS extent is 74mm.

Effectively, this shared*.nii.gz dataset is only about (PT: edited to do the algebra of extents-math correctly now) 1/2 the size of the other one in each dimension (so meaning that one is about one eighth the size of the other), so using the same voxel size, we see that graininess more. If I overlay shared_t1 on MNI (as noted below, I have to fix qform_code and sform_code info and do a center of mass shift for shared_t1), you can see this size comparison:

So, seeing the voxels in this 1mm data is not a GUI problem, it is just a representation of the data as it is. If you would like to make the brain appear smoother, you could do the "define datamode" set of steps I mentioned above. Doing that and asking for cubic interpolation with upsampling to 0.1mm makes the shared dataset look like it does on the right here:

It is a data representation choice, but I prefer to deal with the data in the granularity in which it exists.

Separate issue

The coordinates and header information for this shared dataset are a bit odd. The coordinate origin (x,y,z)=(0,0,0) is not in a central-ish part of the brain (e.g., in the anterior commissure or near the center of mass), it is in fact in the corner of the dataset. If I lock the GUI windows together and jump to the coordinate origin for both the MNI and shared_dset, respectively, the crosshairs show where we are in the data:

This coordinate setup is typically not a desirable property in a reference dataset. This can be fixed/changed/altered, but if there are overlay datasets prepared in the same way, then those much be moved in same way to preserve grid overlap.

Also, if this is a reference space dataset, the NIFTI header information should typically reflect this. This is the dreaded terrain of the qform_code and sform_code. The TL;DR: code values of 3, 4 and 5 denote a reference volume; 1 denotes original space; 2 is the dreaded ambiguous one; and 0 means unknown. The values for the MNI template, a standard space reference, are in the expected range (MNI specifically is 4, by definition):

nifti_tool -disp_hdr -field qform_code -field sform_code -infiles MNI152_2009_template.nii.gz 

N-1 header file 'MNI152_2009_template.nii.gz', num_fields = 2
  name                offset  nvals  values
  ------------------- ------  -----  ------
  qform_code           252      1    4
  sform_code           254      1    4

and those of this other shared_t1 reference are actually just 1, looking like an original space dataset:

nifti_tool -disp_hdr -field qform_code -field sform_code -infiles shared_t1.nii.gz 

N-1 header file 'shared_t1.nii.gz', num_fields = 2
  name                offset  nvals  values
  ------------------- ------  -----  ------
  qform_code           252      1    1
  sform_code           254      1    1

This might be a more minor consideration, but it I just note it as a header feature. This, as well, would have to be updated in any accompanying files if it is altered/changed here.

--pt

[Annen]

Thanks a lot Paul for clearing that up, and showing a way to smooth the image within the viewer.

[rickr]

This isn't too important, but just to clarify a bit, the autobox text:

Extent auto bbox: R=-80.000000 L=-11.000000  A=-99.000000 P=-11.000000  I=2.000000 S=76.000000

seems to show extent sizes of 69, 88 and 74.

It is a little confusing in that, for example, R and L seem to mean rightMOST and leftMOST, presumably so that one can readily distinguish them. But since the R and L signs are both negative, they are both actually right coordinates. Each of 3 of these pairs shares the same sign. That is not the case for the MNI template.

So the grid of the shared dataset seems to be 2mm.
Either way, it would be good to figure out how the 2mm template header got replaced with a 1mm template header.

-rick

[ptaylor]

Hi, Rick-

Thanks for pointing that out about extents, I was careless with the algebra.

--pt