Bad registration with multimodal registration

SCT
1

Hello,

Thank you for the wonderful SCT package, it has been very useful for my research. I am having issues with generating a good registration with sct_register_multimodal for a specific study.

1. A description of the problem. (What are trying to do? What did you expect to happen? What actually happened?)

I am trying to register a diffusion MRI volume to an anatomical T1 scan and I am having issues with a specific study in our data set. The following images show the problem:

DWI Mean segmented with deepseg_sc

Screen Shot 2021-09-26 at 11.29.29 AM
Screen Shot 2021-09-26 at 11.29.29 AM848×577 40.8 KB

T1 volume segmented with deepseg_sc

image
image840×539 41.8 KB

Registered DMRI volume

image
image733×432 21.8 KB

The problem is the waviness of the registered dmri, which causes problems later on in our analysis.

2. Commands and terminal output. (What are the exact commands you used? Please copy and paste the full output of the command from your terminal.)

sct_register_multimodal \
    -i "${REG_DIR}/${WEIGHT}_moved.nii.gz" \
    -iseg "${SEG_DIR}/${WEIGHT}_seg.nii.gz" \
    -d "${DMRI_DIR}/dmri_dwi_mean.nii.gz" \
    -dseg "${SEG_DIR}/dmri_dwi_mean_seg.nii.gz" \
    -ofolder ${REG_DIR} \
    -param step=1,type=seg,algo=centermass:step=2,type=seg,algo=bsplinesyn,metric=MeanSquares,smooth=1,iter=5:step=3,type=seg,algo=slicereg

I have also tried registering anatomical scan to template and registering diffusion to template and using sct_concat_transfo to generate the diffusion->anat transform, but this has yielded similar results.

I can share files via email, to help debug this problem.

Thanks for your help!

2

Hi @vsivan,

Thank you for reaching out.

The registration problem is caused by a problematic spinal cord segmentation, which in turn is caused by poor image quality. This is evidenced by the difficulty to delineate the spinal cord with naked eyes on the sagittal T1w image.

What T1w sequence was that? I am surprised of the low cord/CSF contrast. Also, it says “t1_moved” at the bottom left of the image: is it possible that you performed the segmentation on a registered image?

Julien

3

Hi @jcohenadad ,

Thank you for your quick response.

The imaging sequence for the T1 is MPRAGE Pre contrast.

The image says t1_moved because I use center of mass registration to align the anatomical with the diffusion. This was done because we found that sct_register_multimodal tended to diverge when the field of view between diffusion and anatomical were very different for other studies in our dataset. This divergence does not seem to happen for this particular subject.

I investigated a bit further and found that deepseg does do better if the t1 is not aligned using COM registration (shown below).

T1 image overlaid with segmentation

image
image875×623 79.7 KB

When I try to register the diffusion volume to this unregistered t1 volume, I still get problematic registration artifacts using the exact same sct_register_multimodal parameters (shown below).

Registered diffusion volume

image
image884×617 48.5 KB

Please let me know if you have any ideas for what I can try to remedy the issue.

Thanks again!

4

The image says t1_moved because I use center of mass registration to align the anatomical with the diffusion. This was done because we found that sct_register_multimodal tended to diverge when the field of view between diffusion and anatomical were very different for other studies in our dataset.

I recommend against doing this first pre-registration, because the resolution of the DWI scan is much lower than that of the T1w scan, hence the segmentation is suboptimal. The different FOVs between T1w and DWI scans should not be a problem, because all scans within a session are within the same physical coordinate system (encoded with the qform field of the NIfTI header).

When I try to register the diffusion volume to this unregistered t1 volume, I still get problematic registration artifacts using the exact same sct_register_multimodal parameters (shown below).

You could try to smooth the spinal cord (sct_smooth_spinalcord) and then re-run the segmentation. See example in the tutorial.

5

Hi @jcohenadad

Smoothing did the trick. Thank you so much for your help!

1 Like