Final transform module

[thibaultdvx]

First of all, lots of outdated files deleted: networks/old_network and the associated tests in tests/unittests/nn.

On transforms

Config classes for TorchIO transforms

Supported transforms are:

• RescaleIntensity
• ZNormalization
• Mask
• Clamp
• CropOrPad
• Resize
• Resample
• EnsureShapeMultiple
• Crop
• Pad
• RemapLabels
• OneHot
• NanRemoval (from old clinicadl)

Supported augmentations are:

• OneOf
• RandomFlip
• RandomAffine
• RandomElasticDeformation
• RandomAnisotropy
• RandomMotion
• RandomGhosting
• RandomSpike
• RandomBiasField
• RandomBlur
• RandomNoise
• RandomSwap
• RandomGamma

Transforms object

The main object of the module, Transforms, now supports config classes as inputs (and still supports raw transforms as well). E.g:

>>> from clinicadl.transforms import Transforms, get_transform_config, Patch
>>> import torchio as tio
>>> transforms = Transforms(
        extraction=Patch(),
        image_transforms=[get_transform_config("ZNormalization")],
        sample_transforms=[tio.Resize((16, 16, 16))],
        image_augmentations=[get_transform_config("RandomBlur")],
        sample_augmentations=[tio.RandomAffine()],
    )
>>> transforms.get_transforms()

(Compose([ZNormalization(masking_method=None)]),
 Compose([Resize(target_shape=[16 16 16], image_interpolation=linear, label_interpolation=nearest)]),
 Compose([RandomBlur()]),
 Compose([RandomAffine()]))

The whole transforms module is tested with 100% coverage.

[camillebrianceau]

Just one thing: What do you think about doing something like this for the Subject class? This way, we are forced to have an image and a label in the class, while still being able to work with all the TorchIO transforms.

class ClinicaDLSubject(tio.Subject):
    """
    A TorchIO Subject with an 'image' attribute, a TorchIO ScalarImage and a 'label' attribute, which can be either a float, int or torch.Tensor.
    """
    image: tio.ScalarImage
    label: Union[float, int, torch.Tensor]

Otherwise, LGTM ! Thanks :)

[camillebrianceau]

Suggested change

	self.mask = mask
	self.mask = Path(mask)

[thibaultdvx]

In this case, mask is not a path, but rather a suffix. E.g. with Mask("brain"), the mask will be expected in sub-..._ses-..._brain.nii.gz.
So, I'm not sure we should put it in a Path object.

[thibaultdvx]

Thanks for the review @camillebrianceau. Sounds a good idea, let me check if it works!

[thibaultdvx]

As you suggested @camillebrianceau, I added a subclass of TorchIO Subject with mandatory image and label arguments. I had to overwrite the __copy__ method to make it works with TorchIO's transforms. I'm not a fan of the name ClinicaDLSubject because it is confusing with Subject that we are building in the clinicaio project. So, I suggested the name DataPoint (I was not really inspired haha), but please fell free to suggest if you have a better name.

[thibaultdvx]

I didn't want to put image in the name, as it would be confusing with ImageSample.

[thibaultdvx]

Besides, I added two TorchIO transforms I had forgotten: Resample and ToCanonical.