Add function to mask tissue regions

[LLehner]

Once ready, closes #50

Adds a function to mask an image. Given a user-provided shape, either crop the image such that only pixels within the mask remain or if Invert=False mask the image (only keep pixels outside the shape).

example calls for image masking with user provided shape:

mask_image(sdata, "test", "mask", invert=True)

mask_image(sdata, "test", "mask", invert=False)

Alternatively one can use a threshold + watershed approach to automatically segment and mask the image.
E.g. given the following test image:

mask_image(sdata, "test_2", None, invert=True, automatic_masking=True, threshold=50)

yields

with threshold=80:

For automatic region detection the threshold is image-specific.

Right now the code is still a draft, since i wasn't sure how to properly embed it into your current codebase, so the main questions are:

1. what does the input look like (e.g. instead of spatialdata it will be a project object, but how does it look like when calling the function and how to access the sdata object?).
2. Are the images always grayscale?
3. Should the cropped image be completely resized or will setting masked pixels to 0 suffice?
4. is the mask_filtering module the correct place for this function?
5. Should the automatic approach be more robust and use more novel object detection algorithms or allow for users to plug their own methods in?

[sophiamaedler]

Looks like a great start already! In what form does the user need to provide a shape at the moment? can this be done interactively in Napari?

1. what does the input look like (e.g. instead of spatialdata it will be a project object, but how does it look like when calling the function and how to access the sdata object?).

I would implement this as a method to the project object so that we can e.g. call:
project.mask_input_image(mask, ....)

Under the hood this would then access the spatialdata object that is associated with the project file. You can get it via:
sdata = self.filehandler.get_sdata()

So I guess an implementation along the lines of:

def mask_input_image(mask: ..., invert: bool = False):
    sdata = self.filehandler.get_sdata()
    masked = mask_image(sdata, "test", "mask", invert=True)
    self.filehandler. _write_image_sdata(masked, 
        "input_image",
        channel_names=...,
        scale_factors=....,
        chunks=(1, 1000, 1000),
        overwrite=True)

Currently the _write_image_sdata is still a function belonging to project but it needs to be relocated (and improved) to the filehandler. (the filehandler was introduced more recently when I realized that we otherwise encounter issues with multithreading some operations but the migration has not been completed in full yet).

2. Are the images always grayscale?

So far we have never worked with RGB images. @namsaraeva is working on this for the first time now.
But at the moment I think its fine to implement only for images of the form CXY.

3. Should the cropped image be completely resized or will setting masked pixels to 0 suffice?

Should be completely resized as this will improve run time of all downstream steps. We have some checks where code is not executed if an area only contains 0 pixels, but these checks only go into effect if NO non-zero values are contained.

4. is the mask_filtering module the correct place for this function?

mask_filtering module here refers to operations to filter generated segmentation masks (e.g. to only keep masks with a min cell size). I think the best place for it would be under pipeline/_utils/helper.py.

5. Should the automatic approach be more robust and use more novel object detection algorithms or allow for users to plug their own methods in?

I think it would be nice to have the possibility for users to plug in their own methods, but I think this is a functionality which will only be used by a few more advanced users. So the default as you implemented it now with a thresholding should be sufficient.

[sophiamaedler]

to improve thresholding here a little we could calculate this on a downsampled version of the image which has also been blurred slightly. This should make it a little easier to detect macro structures (like e.g. tissue) somewhat smoothly.