Establishing ground truths on `ms-rennes-mp2rage` dataset

[AcastaPaloma]

[AcastaPaloma]

Monday June 30, 2025

Ran inference on all cases in ms-rennes-mp2rage using model DeepSeg lesion_ms version r20250626 . View spinalcordtoolbox/spinalcordtoolbox#4945

[AcastaPaloma]

@jcohenadad @plbenveniste
Is this the right place to write my progress? I have multiple script files and the output data. Where should I upload that?

[jcohenadad]

Is this the right place to write my progress? I have multiple script files and the output data. Where should I upload that?

yes, I would say this is the right place. I think the script and output data could be uploaded on the git via a PR, but @mguaypaq can confirm.

It would be useful I think to also upload in this issue some QC reports for quick assessment.

[mguaypaq]

Yes, I think:

• This issue is the right place to upload a QC report and discuss the approach.
• A git pull request on our internal server https://data.neuro.polymtl.ca/datasets/ms-rennes-mp2rage would be the right place to put the code and outputs, and discuss the data format.
  • Assuming you're producing labels, your data outputs should live in the derivatives/labels/sub-* folders (if it's not labels, change the /labels/ part of the path to a different name).
  • Normally, the code/ folder at the root of the dataset is for code used to produce/bidsify the root dataset. Instead, your code should probably live in the derivatives/labels/code/ folder.

Previously you only had "read" permissions for the ms-rennes-mp2rage repo. I just gave you "write" permissions, so that you can upload a branches and open pull requests.

[AcastaPaloma]

Thursday July 3, 2025

Ran inference on 20 cases in ms_nmo_beijing
sct_run_batch is troublesome for me unfortunately, see this issue

I also realized that SlicerCART doesn't support the loading of multiple nodes, which is what SCT outputs (the original image + a .seg.nii.gz file). So basically the mask + original image

Currently developing with @SomeoneInParticular a general iterator module discussed during Slicer Week here that will be able to load all the output from SCT models.

Will look into alternatives in the meantime (native Slicer and slow drag-and-drop maybe).
Will upload QC for ms_rennes_mp2rage and ms_nmo_beijing tomorrow after curation.

[AcastaPaloma]

Question

In ms_nmo_beijing, there are cases where the patient has NMO and not MS - I'm assuming it is futile to run sct_deepseg lesion_ms over them? Or could it be a good test for the model either way. @jcohenadad

Thank you!

[jcohenadad]

I'm assuming it is futile to run sct_deepseg lesion_ms over them? Or could it be a good test for the model either way.

Right, you could run the model on NMO patient, if that helps with segmenting those hyperintense lesions. I think the appearance is more diffuse and larger.

[AcastaPaloma]

Monday July 7, 2025

• Reran inference on ms-rennes-mp2rage for correct formatting with derivatives/labels/ and derivatives/labels/code structure
• Will share QC report before the end of the day
• Labels shared here: https://data.neuro.polymtl.ca/datasets/ms-rennes-mp2rage/src/branch/kyw/add-derivatives/derivatives/labels
• Added script1.py and script.sh in derivatives/labels/ to compensate for issues I ran into with sct_run_batch. I directly iterated through a clone of the dataset and placed the output files in the derivatives/labels folder and created an exterior folder called qc_report which I attached down below as a .zip. Is this flow adequate? Any improvement or suggestions? @jcohenadad @mguaypaq

Please let me know if I did anything wrong, thank you so much!

[AcastaPaloma]

Here is a .zip of the QC report over ms-rennes-mp2rage

qc_report.zip

[plbenveniste]

• Added script1.py and script.sh in derivatives/labels/ to compensate for issues I ran into with sct_run_batch. I directly iterated through a clone of the dataset and placed the output files in the derivatives/labels folder and created an exterior folder called qc_report which I attached down below as a .zip. Is this flow adequate? Any improvement or suggestions? @jcohenadad @mguaypaq

Please let me know if I did anything wrong, thank you so much!

It looks great @AcastaPaloma! I just opened PR 2 so that we can comment on the modifications done to the git-annex repo.

Here is a .zip of the QC report over ms-rennes-mp2rage

qc_report.zip

It's pretty hard to see clearly if the lesions are correctly because the QC is not always centered around the spinal cord (as shown in example below). To solve this, you should first segment the spinal cord with sct_deepseg spinalcord. Then when you segment lesion with sct_deepseg lesion_ms use the -qc-seg flag so that the images showed on the QC are centered around the SC.

[AcastaPaloma]

@plbenveniste Thanks for the feedback; I'm assuming I should rerun inferences with your most recent suggestions?