about problem of using antsAtroposN4 to segmentation brain

[JinDJsuper]

Operating system and version

Ubuntu 20.04

CPU architecture

x86_64 (PC, Intel Mac, other Intel/AMD)

ANTs code version

ants-2.5.0

ANTs installation type

Compiled from source

Summary of the problem

When using antsAtroposN4 to segment the brain without using tissue probability maps, the segmented gray matter images show many edge lines in tpm of gm that should not belong to the gray matter.For example, around the medulla oblongata, its outer contour lines appear.

I want to ask why this situation occurs. I haven't seen similar issues when using FSL. Could it be that my parameter selection is problematic? I used the parameters according to the ANTs example and didn't make any changes myself. For now, I prefer not to use tissue probability maps for segmentation.

Actually, I used my own custom-made dog brain TPM (tissue probability map), smoothed with Gaussian, to aid segmentation, and the same issue occurred. However, using the unsmoothed TPM reduces this phenomenon. Therefore, I suspect that segmentation without using the TPM option might also be problematic.

Commands to reproduce the problem.

antsAtroposN4.sh -d 3 -a dogbrain9.nii.gz -x dogbrain9_mask.nii.gz -c 3 -y 2 -y 3 -w 0.25 -o dogbrain9_

Output of the command with verbose output.

-

Data to reproduce the problem

dogbrain9_SegmentationPosteriors2.nii.gz dogbrain9_SegmentationPosteriors1.nii.gz dogbrain9_Segmentation0N4.nii.gz

[ntustison]

Can you post the input as well?

[JinDJsuper]

dogbrain9_brain.nii.gz dogbrain9_mask.nii.gz

Sorry, I went back to look at the FSL results, and they also have edge lines. However, FSL's segmentation performance without TPM is much worse than ANTs, so I think ANTs might be the best choice.

[cookpa]

I see these faint outlines all the time, I think it's because the partial volume voxels on the border between structures can often be close to the mean intensity of another class. The spatial priors and MRF tend to suppress these outlines, but not to zero.

[JinDJsuper]

I see these faint outlines all the time, I think it's because the partial volume voxels on the border between structures can often be close to the mean intensity of another class. The spatial priors and MRF tend to suppress these outlines, but not to zero.

So, with segmentation done this way, can the resulting images be used for voxel-based statistics of gray and white matter? Can they be considered accurate?

[cookpa]

So, with segmentation done this way, can the resulting images be used for voxel-based statistics of gray and white matter? > Can they be considered accurate?

You've identified some areas where they are clearly not accurate, so you have to deal with that. One way would be masking, eg if you're looking at cortical GM, you probably will not include the border of the medulla oblongata in the ROI.

For the rest, you have to look at your data and see if you think it's accurate enough for VBM with the appropriate smoothing.