Tissue segmentation ICBM152 - 2023

1

Hello,

I just tried using the new ICBM template (ICBM 2023) and realized there was something weird with the segmentation in the frontal region (see picture). I am not sure if it's a bug or expected (is it linked to the air in the frontal sinuses ?). Also; it seems that the eye are classified as skin; when i would have thought they would be classified as CSF.

Capture d’écran 2023-02-15 à 21.50.41
Capture d’écran 2023-02-15 à 21.50.411566×1620 237 KB

Was this file manually generated by SPM or cat12 or is it directly coming from the template? The description of ICBM indicates that the template includes T1w,T2w,PDw modalities, and tissue probabilities maps.

Also, was the segmentation from FreeSurfer created using only the T1w or the T1w and T2w ?

Thanks a lot,
Edouard

2

Oops... this looks pretty bad indeed...
This Tissues file was generated with CAT12 r2166.
The surfaces were generated with FreeSurfer 7.3.2, with the T1 only (this is not a realistic T1 and T2)

We could indeed reuse the TPMs from the MNI website (https://www.bic.mni.mcgill.ca/ServicesAtlases/ICBM152NLin2009). However they include only: GM, WM and CSF (+eyes mask).
=> No skull and no skin, which is mostly what we need to fix here.

I am running some tests on SimNIBS4 / CHARM, I could try running it on this volume, and see it I get anything better out of it.

Would you have a better solution?

3

There is few things i would like to test. Could you send me the output folder from FreeSurfer?

Already in the MRI, the contrast doesn't looks very good in the frontal area:

Capture d’écran 2023-02-16 à 13.14.47
Capture d’écran 2023-02-16 à 13.14.471566×1620 319 KB

and it was looking better in the 2019 version :

Capture d’écran 2023-02-16 à 13.15.51
Capture d’écran 2023-02-16 à 13.15.511566×1620 130 KB

I am wondering if it might not be an issue with FreeSurfer normalization.
Here is a more direct comparison of the T1 before and after FreeSurfer:

Capture d’écran 2023-02-16 à 14.09.06
Capture d’écran 2023-02-16 à 14.09.061920×1244 115 KB

4

I haven't kept track of the decision making for using this sharper and darker volume instead of the old one. Maybe the high contrast and sharp separation between elements was helping with some processes. Or something more irrational, like: I was working on a screen with the brightness set to the max...

Looking again at it now, I agree with you, I don't think it makes much sense, it's exactly the same volume but just darker... I don't even recall where it was coming from...
=> I will restore the older contrast

The FreeSurfer segmentation was computed from the "brighter" volume (the one available on the MNI website), nothing to fix on this end.

The CAT12 segmentation, however, was executed on the dark volume, and this is probably the cause of the problems in segmentation. I'm currently re-running it to get the skin+bone masks, and this time I'll use the csf+wm+gm TPMs from the MNI website.
=> I will update the Tissues volume.

Thanks for your constructive feedback :slight_smile:

5

and this time I'll use the csf+wm+gm TPMs from the MNI website.

Now I'm testing this, I'm not sure about it anymore.
The TPMs from the MNI website (left) produce more realistic voxel classifications than the CAT12 segmentation (right), however the CSF layer is not continuous, which might cause other problems.

image
image837×907 32.3 KB

@edelaire @tmedani Which one do you think I should include in the template "ICBM152 2023"?

6

oh ok; i see. I thought it was the output of FreeSurfer. It looks indeed better to separate GM and WM. but maybe not great for skin and CSF.

By curiosity how did you combined the output of CAT12 and the one files from the template ? I guess you did something like putting all the voxel where the probability of skin is larger than 0.5 to skin, then Skull, GM then WM ? if yes, maybe instead looking at each vowel which tissue has the largest prob might be better ?

I am asking that since, it seems that the GM is way more large in this segmentation than in the output of CAT12.

11bf16ec49b2602629ab5dfc1391bc26ef2c36c8
11bf16ec49b2602629ab5dfc1391bc26ef2c36c8837×907 67.6 KB

I am not sure which one is correct; however :slight_smile:

But it might be only detail at the end :slight_smile:

7

By curiosity how did you combined the output of CAT12 and the one files from the template ?

I started from the tissues volume from CAT12, which include all the tissues.
Then I used the volumes from the MNI website to overwrite the values within the brain mask (using the maximum probability at each voxel):

csf = in_mri_bst('ICBM152_2023_TPM/subjectimage_csf.mat');
wm = in_mri_bst('ICBM152_2023_TPM/subjectimage_wm.mat');
gm = in_mri_bst('ICBM152_2023_TPM/subjectimage_gm.mat');
mask = in_mri_bst('ICBM152_2023_TPM/subjectimage_mask.mat');

TissuesFile = 'C:\Work\Protocols\Templates\anat\ICBM152_2023_TPM\subjectimage_tissues_cat12_volatlas.mat';
tissues = in_mri_bst(TissuesFile);

% White: 1
tissues.Cube((mask.Cube == 1) & (wm.Cube >= gm.Cube) & (wm.Cube >= csf.Cube)) = 1;
% Gray: 2
tissues.Cube((mask.Cube == 1) & (gm.Cube >= wm.Cube) & (gm.Cube >= csf.Cube)) = 2;
% CSF: 3
tissues.Cube((mask.Cube == 1) & (csf.Cube >= wm.Cube) & (csf.Cube >= gm.Cube)) = 3;

bst_save(TissuesFile, tissues, 'v7');

I am asking that since, it seems that the GM is way more large in this segmentation than in the output of CAT12.

Indeed, the GM is much thicker when using the MNI ICBM152 TPMs.

But it might be only detail at the end :slight_smile:

In some places there is almost a centimeter of difference for the GM boundaries, so it would have a major impact on any process that uses this tissue classification (example: the automatic labeling of SEEG electrodes).

@Anand_Joshi @John_Mosher
Do you have any opinion on which volume we should use?

1 Like
8

I updated the ICBM152 2023 template:

  1. Replaced the high-contrast/low-brightness T1 with the previous brighter volume.
  2. Re-ran CAT12 2170, and replaced all the volume parcellations
  3. Ran SimNIBS4/CHARM to get a new Tissues labelled volume, including all the new tissues (different types of bones, muscle, blood...)

image
image560×605 59.9 KB

1 Like
9

This looks great !! thanks a lot :slight_smile:

Is there any simple way to remove/merge specific tissue from the segmentation? i guess i can always export to Matlab and change there but I was wondering if there was a way from Brainstorm GUI?

Also, is there a way to run SimNIBS4/CHARM from Brainstorm?

10

Is there any simple way to remove/merge specific tissue from the segmentation? i guess i can always export to Matlab and change there but I was wondering if there was a way from Brainstorm GUI?

No, but we need something like this.
I thought however that this is something that would fit better directly in the functions that need the tissues to be merged, so that different pipelines can keep on using the same FEM mesh in the database without having to duplicate it with different combinations of merge tissues (with risks of not selecting the correct file).
How do you use this file, and what would you like to merge and when?

is there a way to run SimNIBS4/CHARM from Brainstorm?

I'm working on it :slight_smile: