Training segmentation models without the head and the brain stem

[plbenveniste]

In this issue, I explore how removing the brain and the brain stem improves the performance of the model for segmenting spinal lesions in MS.

The brain/brain stem were removed using the contrast agnostic model with sct_deepseg (version: git-master-a866fc666681eca5e7e075b2f6174be0d670f6dd)

The code is currently iterating over every image to create a new msd dataset. The command used was :

python ms-lesion-agnostic/monai/1_create_msd_data_head_cropped.py -pd ~/net/ms-lesion-agnostic/data/ -po ~/net/ms-lesion-agnostic/msd_data/ --lesion-only --canproco-exclude canproco/exclude.yml

Related to #21

[plbenveniste]

I checked if all files had a same directions and found that there was a problem with only one file: sub-P167_UNIT1_desc-rater3_label-lesion_seg.nii.gz.

The problem comes from the origin file which is missing the Sform and qform.

TODO:

• fix the segmentation in the initial dataset and push to the repo
• re-run the cropping of the head for the msd dataset

[plbenveniste]

The origin file was corrected using the following piece of code and pushed to branch plb/fix_p167_lesion_seg:

import os
import nibabel as nib

image_path = "/Users/plbenveniste/tmp_romane/ms_lesion_agnostic/data/basel-mp2rage/sub-P167/anat/sub-P167_UNIT1.nii.gz"
label_path = "/Users/plbenveniste/tmp_romane/ms_lesion_agnostic/data/basel-mp2rage/derivatives/labels/sub-P167/anat/sub-P167_UNIT1_desc-rater3_label-lesion_seg.nii.gz"

image = nib.load(image_path)
label = nib.load(label_path)

# Save the new label with the same header as the image
new_label = nib.Nifti1Image(label.get_fdata(), image.affine, image.header)
nib.save(new_label, "/Users/plbenveniste/tmp_romane/ms_lesion_agnostic/data/basel-mp2rage/derivatives/labels/sub-P167/anat/sub-P167_UNIT1_desc-rater3_label-lesion_seg.nii.gz")

PR is opened and ready for review

[plbenveniste]

The monai model is currently being trained (on koios) with the same parameters as the current SOTA model:

CUDA_VISIBLE_DEVICES=1 python ms-lesion-agnostic/monai/train_monai_unet_lightning.py --config ms-lesion-agnostic/monai/config.yml

The MSD dataset used is: /home/plbenveniste/net/ms-lesion-agnostic/msd_data/dataset_2024-08-13_seed42_lesionOnly.json

[plbenveniste]

The model training and validation curves are displayed below:

It seems that in terms of Dice score, the model didn't outperform our previous SOTA model. However, the validation loss was reduced thanks to the removal of the head and the brain stem.
Looking at the performance of the model on the test set might give us more insights on how it is performing compared to the previous SOTA model. To be done

[plbenveniste]

To compute the performance of this model :

CUDA_VISIBLE_DEVICES=1 python ms-lesion-agnostic/monai/test_model.py --config ms-lesion-agnostic/monai/config_test.yml --data-split test

To compute the figures afterwards:

python ms-lesion-agnostic/monai/plot_performance.py --pred-dir-path ~/net/ms-lesion-agnostic/results_cropped_head/2024-08-13_10\:33\:43.552507/test_set/ --data-json-path ~/net/ms-lesion-agnostic/msd_data/dataset_2024-08-13_seed42_lesionOnly.json --split test

Here are the results: