Optimal Parameters for sct_register_multimodal?

SCT
1

Hello everyone,

I have been working with the sct_register_multimodal tool and experimenting with different parameter settings for image registration between T1-weighted and fMRI data. Specifically, I tried the following configurations:

  1. -param step=1, type=im, algo=dl

    image
    image848×698 33.6 KB

  2. -param step=1, type=seg, algo=slicereg: step=2, type=seg, algo=centermass, slicewise=1, iter=3

    image
    image906×716 33.5 KB

I observed that the second configuration seems to yield better results. However, I wanted to ask the community: Are there any other recommended or optimal parameter settings that might produce even better results than what I’ve tried so far?

Here is my data
fmri_spinal_moco_mean_seg.nii.gz (2.9 KB)
fmri_spinal_moco_mean.nii.gz (3.7 MB)
t1_spinal.nii.gz (6.4 MB)

Thanks!!!

2

Hi,

  • the t1_spinal_seg.nii.gz is missing. I generated it using, but in the future it would be helpful to provide all the files necessary to be able to reproduce your results. I generated it using sct_deepseg_sc (see qc report)
  • Note that slicewise=1, iter=3 are irrelevant with algo=centermass, since it is always slicewise, and there is no optimization
  • Also, running algo=centermass after algo=slicereg makes step=1 irrelevant since the centermass algo overwrite any transformation estimated with slicereg, since it is simply aligning the centermass between the source and destination segmentations
  • I suggest you provide QC reports, instead of static images showing registration results. It is impossible for us to assess the quality of registration on a static image, since we need to be able to toggle between the two images to appreciate the quality of co-registration
  • Note that the fMRI data are strongly hampered by signal dropout caused by susceptibility artifacts, making it difficult to obtain a proper registration
  • It is not clear from your original message if you were using the fMRI data as the target or the source for registration. I used the fMRI as the destination, since it is much faster (ie: less axial slices)
  • In terms of performance, that one looks fine to me:
    sct_register_multimodal -d fmri_spinal_moco_mean.nii.gz -dseg fmri_spinal_moco_mean_seg.nii.gz -i t1_spinal.nii.gz -iseg t1_spinal_seg.nii.gz -param step=1,type=seg,algo=centermass -qc qc

Here is my QC report: qc.zip (941.3 KB)

3

Hi Julien,

Thanks for your detailed explanation and support. Under your guidance, most of the data achieved very good registration results, which has been extremely helpful.

However, I have encountered an issue with a few data. After registering the fMRI data to the template, I noticed a shift, as shown in the image (and in the data).

image
image1920×973 181 KB

Upon further inspection, I found that there is no issue when registering T1 to fMRI, nor is there any problem when registering T1 to the template. The problem arises when applying the transformation from fMRI to the template, which leads to the observed shift.

    sct_apply_transfo -i ${datadir}/${sub}/spinal/fmri_spinal_moco_mean_seg.nii.gz  \
                      -d ${template}  \
                      -w ${datadir}/${sub}/spinal/${base_out}/warp_fmri_spinal_moco_mean2t1_spinal.nii.gz ${datadir}/${sub}/t1/spinal/reg/warp_anat2template.nii.gz\
                      -o ${datadir}/${sub}/spinal/${base_out}/fmri_spinal_moco_mean_seg_template.nii.gz

I have tried different interpolation methods, but it seems that none of them resolve the issue. I would greatly appreciate your suggestions on how to address this problem. Thank you again for your help.

Best,
Xiaomin

4

I was able to reproduce the issue you describe.

For posterity, I have updated the commands to use the exact paths found in the sub.zip “data” link that you shared (to help @jcohenadad debug);

frmi → t1 → template

cd sub/sub  # extracted folder
sct_apply_transfo -i fmri_spinal_moco_mean_seg.nii.gz  \
                  -d ${SCT_DIR}/data/PAM50/template/PAM50_cord.nii.gz  \
                  -w reg_test/warp_fmri_spinal_moco_mean2t1_spinal.nii.gz t1/reg/warp_anat2template.nii.gz\
                  -o fmri_spinal_moco_mean_seg_template.nii.gz

image
image1920×844 179 KB

template → t1 → fmri

cd sub/sub  # extracted folder
sct_apply_transfo -i ${SCT_DIR}/data/PAM50/template/PAM50_cord.nii.gz   \
                  -d fmri_spinal_moco_mean_seg.nii.gz \
                  -w t1/reg/warp_template2anat.nii.gz reg_test/warp_t1_spinal2fmri_spinal_moco_mean.nii.gz \
                  -o PAM50_cord-to-fmri.nii.gz

image
image1920×844 61.4 KB

I will continue to try to debug. I just wanted to quickly share this post to save time with reproducing the issue.

Kind regards,
Joshua

5

Thank you so much Joshua!

6

Apologies for the late reply! I had a little more time to look at this now. :slight_smile:

Focusing in on just the “fMRI to T1” registration, I see the same shift when warping the fMRI image to the T1 space by running the following command:

sct_apply_transfo -i fmri_spinal_moco_mean_seg.nii.gz  \
                  -d t1/t1_spinal.nii.gz  \
                  -w reg_test/warp_fmri_spinal_moco_mean2t1_spinal.nii.gz \
                  -o fmri_spinal_moco_mean_seg_t1_spinal.nii.gz

image
image1913×854 144 KB

When I double-check the fMRI-T1 registration results by running the same command, just on the anat image instead of the segmentation, it’s fine?

sct_apply_transfo -i fmri_spinal_moco_mean.nii.gz  \
                  -d t1/t1_spinal.nii.gz  \
                  -w reg_test_2/warp_fmri_spinal_moco_mean2t1_spinal.nii.gz \
                  -o fmri_spinal_moco_mean_seg_t1_spinal.nii.gz

image
image1913×854 159 KB

So, if the anat image can be warped OK, then something seems to be wrong with the _seg image specifically.

When I check the headers for both the original and the _seg image, I notice that the sform of the segmentation is shifted!

image
image1335×345 64 KB

However, even though the sform is shifted, the segmentation is aligned with the anat image:

image
image1920×851 52.5 KB

And, likewise, if I “fix” the sform of the segmentation (so that they are in the same space) by running the command sct_image -set-sform-to-qform -i fmri_spinal_moco_mean_seg.nii.gz, the shift appears!

image
image1920×851 58 KB

So, tl;dr:

  • The segmentation is not in the same space as the anat image (due to different sforms).
  • Fixing the sform (but leaving the data array as-is) doesn’t fix the problem, since both the sform and the data array have this shift.

@Xiaomin_Lin, how was the segmentation generated?

7

Hi Joshua,

I just wanted to sincerely thank you for your help with the debugging! I took a look using Nibabel and, sure enough, I found that the data with the shift have discrepancies between their _seg sform and the original sform. I’m still not entirely sure how this happened, but I suspect it might be related to some weirdness that occurred when I manually modified the SEG file.

In any case, I really appreciate your assistance. Without your suggestion, I would never have thought to check this! Thanks again for pointing me in the right direction.

Best,
Xiaomin

1 Like
8

Thank you for such kind words! I’m very happy to help, and I appreciate your continued participation on the forum. :hearts:

I’m glad it helped! For future reference, you can use the command sct_image -header to check this information. (We created this flag as a mirror of the “fslhd” utility, so that users wouldn’t have to install the entirety of FSL just to check the header metadata. It’s a very useful tool!)

Anyway, feel free to continue to open new forum posts any time you have questions or concerns, and we’ll be happy to assist. :slight_smile:

Kind regards,
Joshua