blurry multimodal template output with antsMultivariateTemplateConstruction2

[gjoue]

Hi,

Can you help me troubleshoot what I'm doing wrong when trying to create a multimodal study template using antsMultivariateTemplateConstruction2. The final template is fuzzy as if none of the input images were aligned properly.

Inputs are pairs of high-resolution slab T2w-TSE images and 1-mm isotropic whole-brain T1w-MPRAGE images that had been upsampled to the T2w-TSE resolution. I had aligned each subject's N3-corrected and skull-stripped T2w image to the upsampled, N3-corrected and skull-stripped T1w with antsRegistration (finally got slab to whole brain reg to work, thanks for your help. To my eye, the pairs of images seem aligned. Here is the T2w as a red outline overlaid on the corresponding T1w. I also checked the actual volumes side by side:

)

This is the antsMultivariateTemplateConstruction2 command I used:

antsMultivariateTemplateConstruction2.sh \
	   -d 3 \
	   -o ${p_OUT}/${mod}_ \
	   -r 1 \
	   -n 0 \
	   -w 1x1 \
	   -k 2 \
	   -f 6x4x2x1 \
	   -s 2x1x0x0vox \
	   -i 4 \
	   -c 2 \
	   -j 8 \
	   -m MI \
	   -t BSplineSyN \
	   $T1MPRAGE_T2res $T2 >> $log 2> $logerr

The final template output looks like this:

Thanks for any pointers!
Gina

[cookpa]

It looks like your input is in the wrong order. It needs to be ordered by subject, not all the T1w followed by all the T2.

You can instead use a CSV file eg

t1_sub1.nii.gz,t2_sub1.nii.gz
t1_sub2.nii.gz,t2_sub2.nii.gz
...

Then just pass my_subjects.csv instead of the individual images.

Beware that there was a problem reading CSV files from Excel / Python, which write Windows newlines even on non-Windows platforms. This has been fixed in the latest ANTs.

[gjoue]

Hi Philip,

That was exactly it! I made sure the different modalities were grouped/ordered by subject. The template looks much better, but it seems too blurry (we want to use it for atlas auto-segmentation)?

e
I assume template images can get sharper than that, and it's a matter of adjusting the parameters. Would you have some advice on what could be promising things to try or an approach to test? I'm using ANTS v.2.5 in linux, and this was the command (based on the TemplateBuildingExample and what I read in the ANTs forums):

antsMultivariateTemplateConstruction2
-d 3
-o ${OUT}
-r 1
-n 0
-w 1x1
-k 2
-f 6x4x2x1
-s 2x1x0x0vox
-i 4
-c 2
-j 8
-m MI
-t BSplineSyN
${fs_IN[@]}

Thank you very much for any pointers!
g

[gjoue]

Hoping that everyone's back and refreshed from summer break...I'm still stuck on this one...Does anyone have any suggestions?

[ntustison]

Although definitely on the lower quality side of the various templates I've seen made with ANTs, I wouldn't say that your results are indicative of failure or having done something wrong. It very well could be the quality of data, specifically one of your modalities. For example, if you're coupling high-quality, isotropic T1s with anisotropic T2s, that's definitely going to effect how well your T1s are aligned (especially if they're weighted equally, i.e., ``-w 1x1`). You might want to take a step back and start with just a T1 template as a baseline and then slowly add the contribution of the T2s by playing with the weight and seeing what effect that has on the quality of your template.

[ntustison]

Also, I tend to favor keeping the skull whenever possible during template construction. So after aligning the T2 with the skull-stripped T1, you might want to throw the whole head T1 back into the multivariate template construction.

[cookpa]

Inputs are pairs of high-resolution slab T2w-TSE images and 1-mm isotropic whole-brain T1w-MPRAGE images that had been upsampled to the T2w-TSE resolution.

What is the resolution of the T2w-TSE? Looks like it has thick slices and you have downsampled the T1w in the slice direction.

[gjoue]

Thank you very much, Nick and Philip, for responding and your advice!

Nick, what would you recommend as the minimum size for a toy dataset to try out different weights on the T2w-TSE (we have 100+ subjects and it took 448h on 8 cores)?

Yes, you are right, Philip. Technically, I didn't totally upsample the T1w-MPRAGE, because our T2w-TSE resolution is 0.4 x 0.4 x 2. How did you see that the slices were thick/What were the cues that led you to conclude that it was downsampled?

We are interested in using the group template for atlas-based automatic segmentation. We are interested in mostly the medial temporal lobe and limbic regions. So would you guys recommend just playing with the weights, and not take the blurriness as a problem? I actually had done group templates for each modality separately, too. Here they are in case helps with the discussion.

The T1-only group template:

The T2-only group template:

Thank you!

[ntustison]

I would just start with ~10.

That could point to another problem---you should check the alignment of all your input subjects to the template as misalignments will contribute to the bluriness.

Finally, for anisotropic data, such as these, I like to use the anisotropic MRI super-resolution tool available in ANTsPyNet/ANTsRNet. Complete example here.