DWI mean SC seg

[valosekj]

Currently, I use the contrast-agnostic model to segment mean DWI image:

https://github.com/valosekj/dcm-brno/blob/main/02_processing_scripts/02_process_data.sh#L347-L348

This SC seg is consequently used to bring the template to DWI space:

https://github.com/valosekj/dcm-brno/blob/main/02_processing_scripts/02_process_data.sh#L353-L359

However, when debugging #18, I noticed that the contrast-agnostic segmentation is slightly shifted. So I tried sct_deepseg_sc and found out that it might actually provide a better segmentation:

gif

TODO: compare sct_deepseg_sc and contrast-agnostic on more subjects

[valosekj]

TODO: compare sct_deepseg_sc and contrast-agnostic on more subjects

Okay, tested. The contrast-agnostic seg is shifted at the superior part of the FOV for almost all subjects. --> switching to sct_deepseg_sc in 6787d06

[valosekj]

Resolved in 6787d06

[valosekj]

I also tried contrast-agnostic v2.3, based on the recommendation from @naga-karthik:

If possible, if you could try using the v2.3 model instead? The rationale is that between models v2.3 and v2.4, I added the canproco dataset and I remember that canproco had a lot of shifted segmentations in the GT.

Link to the data used.

conda activate monai
~/code/contrast-agnostic-softseg-spinalcord

# v2.3
python monai/run_inference_single_image.py --path-img /Users/valosek/Downloads/sub-1860B6472B/dwi/sub-1860B6472B_ses-1860B_acq-ZOOMit_dir-AP_dwi_crop_crop_moco_dwi_mean.nii.gz --path-out /Users/valosek/Downloads/sub-1860B6472B/dwi/contrast_agnostic_v2.3 --chkp-path /Users/valosek/Downloads/model_soft_bin_20240410-1136

# v2.4
python monai/run_inference_single_image.py --path-img /Users/valosek/Downloads/sub-1860B6472B/dwi/sub-1860B6472B_ses-1860B_acq-ZOOMit_dir-AP_dwi_crop_crop_moco_dwi_mean.nii.gz --path-out /Users/valosek/Downloads/sub-1860B6472B/dwi/contrast_agnostic_v2.4 --chkp-path /Users/valosek/Downloads/nnunet_seed=50_ndata=7_ncont=9_pad=zero_nf=32_opt=adam_lr=0.001_AdapW_bs=2_20240425-170840/

The shift in the top slices is presented for all versions (v2.3, v2.3, SCT):

gif

[naga-karthik]

Hey Jan! Before I start debugging this issue a bit deeper could you try the edge padding option when running inference from the model v2.4/v2.3 downloaded from the contrast-agnostic repository (i.e. not using SCT for inference)

Basically from the command you posted above, the change would be:

python monai/run_inference_single_image.py --path-img /Users/valosek/Downloads/sub-1860B6472B/dwi/sub-1860B6472B_ses-1860B_acq-ZOOMit_dir-AP_dwi_crop_crop_moco_dwi_mean.nii.gz --path-out /Users/valosek/Downloads/sub-1860B6472B/dwi/contrast_agnostic_v2.4 --chkp-path /Users/valosek/Downloads/nnunet_seed=50_ndata=7_ncont=9_pad=zero_nf=32_opt=adam_lr=0.001_AdapW_bs=2_20240425-170840/ --pad-mode edge

Instead of zero padding this option does edge padding and in my internal experiments I have seen that this works slightly better for the top/bottom slices. Let me know how it goes!

[valosekj]

Hey Naga! Thanks for the tip! I tried --pad-mode edge, but the predictions are pretty much the same (the shift is still there).

v2.3

v2.4

[valosekj]

Testing the contrast-agnostic model v2.4 after fixing the edge padding as part of SCT v6.4 on DWI mean images. I would say that the segmentations look relatively reasonable now! (of course, some minor manual corrections, especially at the compression levels, would be appropriate)

@naga-karthik, @sandrinebedard, @jcohenadad, what do you think?

[naga-karthik]

It seems that there are no major issues with the contrast-agnostic segmenations, but I notice that some slices are very slightly undersegmented. I am not sure if that's expected with DWI images because the cord/csf boundary is not that clear.

When you tried sct_deepseg_sc, was it convincingly better than the contrast-agnostic model?

[valosekj]

but I notice that some slices are very slightly undersegmented

Yes, you're right! I have the same feeling. I will perform manual correction of these segmentaions. Then we could use them for another iteration of the contrast-agnostic training. @naga-karthik, what do you think?

When you tried sct_deepseg_sc, was it convincingly better than the contrast-agnostic model?

Here is sct_deepseg_sc on the same subjects. Notice that sct_deepseg_sc undersegments even more and struggles with capturing the compressed cord. I would say that the fixed contrast-agnostic model is now consistently better!

[naga-karthik]

Indeed! sct_deepseg_sc is worse than contrast-agnostic for these subjects!

I will perform manual correction of these segmentaions. Then we could use them for another iteration of the contrast-agnostic training. @naga-karthik, what do you think?

Agreed, good plan!