Continuing refactoring of the extraction objects

[thibaultdvx]

Following #685, I continued the refactoring of the extraction classes Image, ROI, Patch and Slice. I suggest to get rid of some complicated features while adding some useful ones:

• For all: remove use_uncropped_image that should be put in Preprocessing.

• Slice: in addition to discarded_slices that only enable slice filtering, I suggest to introduce two new arguments. slices enables slice selection (e.g. Slice(slices=[128, 129, 130])); and borders enables slice filtering at the edge of the image. For example, Slice(discarded_slices=[42], borders=5) will filter out the slice with index 42, as well as the first five and the last five slices in each direction. borders can also be specified for each border (e.g. Slice(borders=((1, 1), (2, 3), (1, 1))).

• Patch: I suggest to enable anisotropic sliding window by allowing the user to pass anisotropic values for patch_size and stride (e.g. Patch(patch_size=(18, 16, 17), stride=(1, 1, 2)).

• ROI: to replace roi_list, roi_mask_location, roi_crop_input, roi_crop_output, roi_template, roi_mask_pattern, roi_custom_template and roi_custom_mask_pattern, I would suggest to have only two arguments. masks, where a list of masks can be passed via their file location (e.g. ROI(masks=["masks/leftHippocampus.nii.gz", "masks/rightHippocampus.nii.gz"])); and crop that enables to filter out regions that are always 0 for all masks, in order to reduce the size of the images.
  In doing so, I purposely release some constraints on the masks that can be used, that I felt unnecessary.

Besides, I also made some minor changes:

• Renaming BaseExtraction to Extraction, in order not to have Base... in our typings;
• Gathering roi, patch, slice and image under the generic word sample. I changed the function names and dosctrings accordingly;
• Uniformizing the docstring between the four child classes (there is still work the be done).

As of now, the classes are not tested. i'm waiting for your opinion first.

[thibaultdvx]

The classes are now tested with 100% coverage.

[thibaultdvx]

I just added some useful class methods to manipulate TorchIO's class Subject. This is crucial as our transforms will manipulate Subject objects. More precisely:

• get_tio_image: puts the image and the label in a Subject, as well as other potential masks that will be used to compute the transforms.
• extract_tio_sample: does the extraction like extract_sample, but works on Subject. All the images in the Subject object will undergo the extraction. For example, in the case of segmentation, extraction will be performed also on the segmentation map.
• format_output: takes a Subject and some metadata and returns a Sample, which is typically the output of the __getitem__ in our CapsDataset. Sample is similar to the old CapsDatasetSample, but it is different depending on the extraction method.

So, in the __getitem__ of our CapsDataset, I imagine it this way:

    participant, session, img_index, sample_index = self._get_meta_data(idx)
    image, image_path = self._get_image(img_index)
    label = self._get_label(img_index)

    tio_image = self.transforms.extraction.get_tio_image(image, label)

    tio_image = self.transforms.image_transforms(tio_image)

    if not self.eval_mode:
        tio_image = self.transforms.image_augmentation(tio_image)

    tio_sample = self.transforms.extraction.extract_tio_sample(tio_image, sample_index)

    tio_sample = self.transforms.sample_transforms(tio_sample)

    if not self.eval_mode:
        tio_sample = self.transforms.sample_augmentation(tio_sample)

    return self.transforms.extraction.format_output(
        tio_sample,
        participant_id=participant,
        session_id=session,
        image_path=image_path,
    )

[thibaultdvx]

In Slice, I removed the argument slice_mode, as duplicating channels can easily be done with a basic transform.

[thibaultdvx]

As discussed in developer meeting, I removed the ROI extraction as it can be done with common transforms (e.g. Mask, potentially followed by Crop).

[camillebrianceau]

LGTM !