Overview of pathology naming

[drbeh]

Create and overview of different pathology components and their naming structure.

[drbeh]

MONAI Pathology-specific Components

Type	Subtype	Component	Backend	Transition Path
data	datasets	PatchWSIDataset	-	It can be extended to use any derivative of ImageReader class; however, some considerations are required to avoid unnecessary caching for backend libraries with intrinsic cache (like OpenSlide). Also reading a subregion can be generalized to include other image types and readers.
data	datasets	SmartCachePatchWSIDataset	-	With a generalized PatchWSIDataset, this can be removed.
data	datasets	MaskedInferenceWSIDataset	-	This can be generalized with a sampling on a low-resolution binary mask, which could work with any datasets and dataloader. It is based on the common concept of image level in digital pathology but that can be abstracted to create a uniform API.
data	image_reader	WSIReader	OpenSlide, cucim.CuImage, TiffFile	Already in MONAI Core
handlers		ProbMapProducer	numpy.ndarray	It has room to improvements like multi-GPU support; however, making it generalizable to non-pathology images requires many changes and additional helper classes.
transforms	spatial	SplitOnGrid	numpy.ndarray, torch.Tensor	This SplitOnGrid and TileOnGrid have very much in common; however, merging them is not trivial since they have different implementations with different efficiencies. Also the combination make the API more complicated with so many options and situations to handle. Need further evaluation of combining approach. Otherwise, we should make the differences more clear and separate their functionalities as much as possible.
transforms	spatial	TileOnGrid	numpy.ndarray
transforms	color			In pathology, color deals with stains, so we can have either transforms/color or transforms/stain; however, to make it fit to a broader biomedical imaging context, color might be a better choice. Also in case we add any transform that manipulate color profiles.
transforms	color	HEStainExtractor	numpy.ndarray	It is worth to explore the posiibility of having the same API as scikit-image or even moving this component there. Efficiency needs to be taken into account for this components.
transforms	color	StainNormalizer	numpy.ndarray	It can be generalized to any dimensional color image with three channel color
transforms	utility	CuCIM	cuCIM transforms	Already in MONAI Core
transforms	utility	RandCuCIM	randomized cuCIM transforms	Already in MONAI Core
metrics	FROC	LesionFROC	numpy.ndarray, torch.Tensor	Specific to 2D color images but there is no fundamental barrier to make it 3D
metrics	FROC	compute_fp_tp_probs	numpy.ndarray, torch.Tensor	Already in MONAI Core
metrics	FROC	compute_froc_curve_data	numpy.ndarray, torch.Tensor	Already in MONAI Core
metrics	FROC	compute_froc_score	numpy.ndarray, torch.Tensor	Already in MONAI Core
networks	nets	TorchVisionFCModel	-	Already in MONAI Core

[wyli]

Thanks for the summary, would be great to specify the proposed source code file location and namespaces. For example, for monai's LMDBDataset module, it is in

monai/data/dataset.py

and it is accessible via

from monai.data import LMDBDataset
from monai.data.dataset import LMDBDataset

do you envision any alternative submodule aliases for the same module? e.g. monai.data.wsi.PatchDataset and where to put the actual files, it seems monai/data/dataset.py is quite lengthy...? cc @ericspod

[ericspod]

Adding more subdirectories is the way to go to organise and make explicit what components they have, so monai/data/wsi and monai/transforms/color and others make sense. Many of our existing transform files may do with reorganisation along thematic lines like this anyway so it won't be out of place.

[drbeh]

@wyli @ericspod, I totally agree with you. monai/transforms/color sounds good to me but let me check on monai/data/wsi to see if there is a way to abstract away WSI part and make it patch-based datasets monai/data/patch. I will check and keep you posted.

[drbeh]

@wyli @ericspod, on a separate note, since there are many interconnected changes, does it make sense to create a new branch for pathology (then merge it after changes are done)?

[drbeh]

@wyli @ericspod, I totally agree with you. monai/transforms/color sounds good to me but let me check on monai/data/wsi to see if there is a way to abstract away WSI part and make it patch-based datasets monai/data/patch. I will check and keep you posted.

I evaluated the components and realized the specific needs of pathology datasets are more tied to the whole slide image rather than being patch based, so I think monai/data/wsi/datasets.py an appropriate place to host these datasets. Although patches can be used in radiology workflows too, here we are dealing with patches from images that cannot be loaded into memory. I will preparing a diagram of the pathology component structure to have a better overall view.

[drbeh]

Refer to #4005 for more details.