Corrupted lesion segmentation

[plbenveniste]

I have identified the following issue in the data used for testing the model:
(in green the manual segmentation/ in red the segmentation used for training/testing the model)

Possible causes for this:

• label reorientation:

  ms-lesion-agnostic/nnunet/convert_msd_to_nnunet_reorient.py

  Line 141 in 718d6a9

  assert os.system(f"sct_image -i {label_file_nnunet} -setorient RPI -o {label_file_nnunet}") ==0

• label resampling to match image dimension:

  ms-lesion-agnostic/nnunet/convert_msd_to_nnunet_reorient.py

  Line 148 in 718d6a9

  assert os.system(f"sct_register_multimodal -i {str(label_file_nnunet)} -d {str(image_file_nnunet)} -identity 1 -o {str(label_file_nnunet)} -owarp file_to_delete.nii.gz -owarpinv file_to_delete_2.nii.gz ") ==0

• label binarization:

  ms-lesion-agnostic/nnunet/convert_msd_to_nnunet_reorient.py

  Line 155 in 718d6a9

  assert os.system(f"sct_maths -i {str(label_file_nnunet)} -bin 0. -o {str(label_file_nnunet)}") ==0

This is a BIG issue!

[plbenveniste]

The problems come from the fact that in the registration step: everything is done using -x linear instead of doing -x nn.

This issue could be spread everywhere in the dataset so maybe it would be wise to fix this before publishing. @jcohenadad Does this mean I should retrain the model from scratch?

[plbenveniste]

Thankfully, I just realized that there should only be an issue for the test images. This is because binarization was done with -bin 0.5 for the train images and -bin 0. for the test images. This is very likely a typo.
This is good news as it means that I "only" need to re-do the evaluations. Not everything!

[plbenveniste]

Fixed in commit: dfb7227

[plbenveniste]

Running this to get a new version of the test set:

python nnunet/convert_msd_to_nnunet_reorient.py --input  ~/net/ms-lesion-agnostic/msd_data/dataset_2025-04-15_seed42.json --output ~/net/ms-lesion-agnostic/postprocessing/output_nnunet_dataset --tasknumber 902

(I commented the part which does the training set to only focus on the test set)

[plbenveniste]

The labelsTs folder was replaced with the updated version. Dataset is still stored here: /home/plbenveniste/net/ms-lesion-agnostic/nnunet_experiments/nnUNet_raw/Dataset902_msLesionAgnostic. Currently updating the performance metrics on the correct label set