Region-based nnUNet training

[plbenveniste]

In this issue, I document my experiments in training and testing a region-based nnUNet.
The model takes into input a Nifti image and outputs the spinal cord (value=1) and the MS lesions (value=2).
The fact that it is region-based means that lesions must be in the spinal cord.

The necessary steps to take are :

• Segment SC for images which have lesion seg but no SC seg (with SCT v6.2)
• Manually correct the SC segs
• Build an nnUNet dataset
• Train a 2D region-based nnUNet (with normal loss and no smoothing loss)
• Train a 3D region-based nnUNet (with normal loss and no smoothing loss)
• Test model on test set
• Compare model performances

[plbenveniste]

In the CanProCo dataset as well as in the stc-testing-large dataset, some subjects were missing a SC segmentation.

• For the 33 CanProCo subjects, the SC was segmented using the region-based model trained on CanProCo (release).
  They were then manually corrected and pushed to CanProCo on branch plb/add_sc_seg_for_regionbased
• For the 186 sct-testing-large subjects, the SC was segmented using the contrast-agnostic-sc-seg model from SCT v6.2. They were then manually corrected and pushed to sct-testing-large on branch plb/add_sc_seg_for_regionbased