Refactor `sct_deepseg` to store model orientation in model dict about spinalcordtoolbox

[NathanMolinier]

code like this one is supposed to take care of this, no?

Yes it is indeed the case, however I don't know if nnUNet is performing any reorientations beforehand.

Also, I wanted to remind that conventions may be different depending on the tools that were used: for example, reorientation could be done before the training as pre-processing using SCT for example or using MONAI's transform.

[joshuacwnewton]

Couldn't we encode the reorientation step directly from MONAI's data preparation phase, while running the inference? Eg: code like this one is supposed to take care of this, no?

Yes, SCT contains this exact same line in our current contrast agnostic code.

MONAI (inference transforms):

spinalcordtoolbox/spinalcordtoolbox/deepseg/monai.py

Line 166 in fef1bea

Orientationd(keys=["image"], axcodes="RPI"),

nnUNetv2 (manually, since we predict directly on a numpy array):

spinalcordtoolbox/spinalcordtoolbox/deepseg/inference.py

Lines 198 to 202 in fef1bea

	# Get the orientation used by the model
	if "SCI" in predictor.plans_manager.dataset_name:
	model_orientation = "RPI"
	else:
	model_orientation = "LPI"

Because we are hardcoding the axcodes inside the inference functions, I wanted to just do a simple reorganization to move these hardcoded values alongside the models themselves.

(This came up because the canproco nnunetv2 model uses the AIL orientation, which required a change to the hardcoded values.)

[jcohenadad]

Perhaps it would be best to add "orientation" to the MODELS dict for each of the nnunet/monai models, then specify LPI/RPI/AIL on a per-model basis, then fetch this value from the model dict when reorienting the input image?

Couldn't we encode the reorientation step directly from MONAI's data preparation phase, while running the inference? Eg: code like this one is supposed to take care of this, no? Tagging @naga-karthik @plbenveniste @Nilser3 @NathanMolinier

[NathanMolinier]

I did not implement any model yet to SCT, but aren't we calling model specific inference scripts when calling the models ?