Lesion segmentation analysis

[plbenveniste]

In this issue, I use the segmentation of the lesions on the spinal cord after cropping the images (issue #37 ).

The analysis of the segmentations accounts for the following:

• Identifying a lesion across slides
• Measuring its volume
• Measuring its center
• Registration between images and segmentations at different time points
• Comparison of the lesions at different time points
• Creating a text file to sum up evolutions of the MS lesions

[plbenveniste]

I worked on a segmentation that was inferred by nnU-Net (the one trained on 1000 epochs) on image sub-cal072_ses-M12_STIR. On this volume, I performed:

• Automatic clustering of the segmented voxels in lesions (the number of lesion is automatically detected) across slices
• Computation of the lesion volumes
• Computation of the center of the lesions

Output:

It outputs a file where each segmented voxel is indexed per lesion (1 for lesion 1, 2 for lesion 2, etc.)

Number of lesions: 4
Volume and center of each lesion (mm3):
Lesion 1 : volume: 42.63 mm3, center: [155.5862069 224.79310345 4. ]
Lesion 2 : volume: 232.26 mm3, center: [159.92405063 170.44303797 4.34810127]
Lesion 3 : volume: 29.4 mm3, center: [164.4 132.3 3. ]
Lesion 4 : volume: 29.4 mm3, center: [218.75 21.5 4. ]

[valosekj]

• Automatic clustering of the segmented voxels in lesions (the number of lesion is automatically detected) across slices
• Computation of the lesion volumes
• Computation of the center of the lesions

Cool! For this, you are using the lesion_seg_analysis.py script from plb/nnunet branch, right?

Be aware that SCT contains sct_analyze_lesion.py function, which does pretty similar things:

$ sct_analyze_lesion -m sub-cal080_ses-M0_STIR_lesion-manual.nii.gz 

...

Lesion count = 1

Measures on lesion #1...
  Volume : 52.92 mm^3

...

Averaged measures...
  Volume [mm^3] = 52.92+/-0.0
  (S-I) Length [mm] = nan+/-nan
  Equivalent Diameter [mm] = nan+/-nan

Total volume = 52.92 mm^3
Lesion count = 1

...

(sct_analyze_lesion.py does not print the lesion center, but this could be easily added)

[plbenveniste]

Currently performing registration between M0 and M12 images using:
sct_register_multimodal -i sub-cal072_ses-M0_STIR.nii.gz -d sub-cal072_ses-M12_STIR.nii.gz -iseg sub-cal072_ses-M0_STIR_seg.nii.gz -dseg sub-cal072_ses-M12_STIR_seg.nii.gz with the segmentations being the spinal cord segmentation (computed with sct_deepseg_sc) used to improve the registration.

The registration is not of great quality because of the very limited number of slices on the z-axis (9 slices in this case). This will surely have a clear impact on the study of the evolution of spinal cord lesions.

[jcohenadad]

what is the purpose of the registration? if it is to identify the same lesions across time points, then it should be fine (given that lesions span several voxels)

[plbenveniste]

The objective is to register images from M0 and M12 to perform comparison of the sc lesions. Yes, it means identifying the lesions across time points.

[valosekj]

Currently performing registration between M0 and M12 images using: sct_register_multimodal -i sub-cal072_ses-M0_STIR.nii.gz -d sub-cal072_ses-M12_STIR.nii.gz -iseg sub-cal072_ses-M0_STIR_seg.nii.gz -dseg sub-cal072_ses-M12_STIR_seg.nii.gz with the segmentations being the spinal cord segmentation (computed with sct_deepseg_sc) used to improve the registration.

Another possible approach might be using sct_straighten_spinalcord; see point 2 in neuropoly/idea-projects#17 (comment). However, sct_straighten_spinalcord requires, besides SC seg, also disc labels (which are not currently available for STIR/PSIR).

[jcohenadad]

from @valosekj:

Another possible approach might be using sct_straighten_spinalcord; see point 2 in neuropoly/idea-projects#17 (comment). However, sct_straighten_spinalcord requires, besides SC seg, also disc labels (which are not currently available for STIR/PSIR).

That's a good idea, although it might be an overkill (straightening takes time). But for visualisation purpose, that could be interesting. In any case, if @plbenveniste needs to have some reasonable level of alignment between time points, he will definitely need disc levels (otherwise, how can he make sure that the coverage is exactly the same across time points).

from @plbenveniste:

The objective is to register images from M0 and M12 to perform comparison of the sc lesions. Yes, it means identifying the lesions across time points.

In that case, you don't need sub-millimetric precision, given that lesions are 'big blobs'. 2-3mm precision in the registration will make it possible to identify the same lesion across time.

[plbenveniste]

After registration, there is still one problem. Even though registration is well performed on slices, because of the anisotropic nature of the image, we can see slices with spinal cord "registered" to slices without spinal cord. That happens, even though both images are well aligned on the x and y axis.

The resulting problem is visible when performing registration on the lesion segmentation images. The time points can't really be compared since they don't superpose well.

Left image = 1st time point image / Right image = 2nd time point image registered to 1st time point
Red labels = 1st time point lesions / Blue labels = 2nd time point labels (registered to 1st time point)

Solution:

• One solution is to perform lesion comparison from M0 to M12 without doing registration, hoping that lesions from the first time point are "similar enough" to those at the second time point and close enough.
• Another solution I want to try is to create a .nii.gz image in which 0=spinal cord and 1=lesion segmentation. From this, I'll directly perform registration to superpose lesions and then compare them.

[jcohenadad]

I'm not sure I fully understand #38 (comment). A few comments:

• you mentioned "the registration is well performed on slices", but the first GIF animation suggests that the two time points are not well registered (notice a shift of ~5cm along the rostro-caudal axis)
• also, what do you mean by registration performed "on slices"? How was the registration done by the way (linking to your code would be useful)
• "both images are well aligned on the x and y axis": in medical imaging, I would recommend to instead use the terms AP, RL, IS to avoid confusion (because an "x" axis can be along AP, RL or IS depending on the orientation of the FOV during acquisition)
• can you make the images brighter/more exposure? it is very difficult to properly see these images
• what are the red and blue labels on the GIF images?

[plbenveniste]

The code is here

[jcohenadad]

additional comments:

• always use QC report
• make sure to add params in your registration, notably: use the seg as step1 (with centermass or slicereg-- play with polynom order), with syn (step2) make sure to use deformation along 1x1x1, metrics: should NOT be meansquare (bc images aquired at diff time points) but instead CC, experience with smoothing (useful in noisy datasets), make sure to use -x linear (or sync)
• you could potentially use the vert labels (but start without-- i think you should be fine without)
• you might be able to register without the segmentation, using the "dl" algo, and if that doesn't work, start with translation and also try syn with various shrink factor.

[plbenveniste]

Changing the parameters improved the registration. The final parameters are:

step=1,type=seg,algo=slicereg,poly=5:step=2,type=im,algo=syn,metric=MI,deformation=1x1x1,smooth=3

I played with the polynomial order, the metric (MI or CC), and the smoothing.

However, even though the spinal cord is well aligned, there is still a problem with the alignment of vertebral labels as we can see in the following:

We can see that the first time point is not overlapping correctly with the second time point, especially when looking at the vertebral levels (look at the green cursor) or when looking at the label from the first time point on both the image from the first time point and the image from the second time point.

I am now trying to add vertebral levels to improve the quality.

[plbenveniste]

I can't manage to use vertebral labels in the following command line.

The error I get is the following:

[valosekj]

Exception: Error: number of source and destination landmarks are not the same, so landmarks cannot be paired.

Do the -ilabel and -dlabel have the same number of labels?