Allowing tile to extend beyond image region by padding. about tiffslide

[One-sixth]

Hey everyone,

It does make sense to support this.

The implementation would be limited to these few lines:

https://github.com/bayer-science-for-a-better-life/tiffslide/blob/56b5bbda36a2792a93b3eabd30c590f004835b41/tiffslide/tiffslide.py#L358-L408

TODOs:

• add keyword only padding argument defaulting to True
• calculate padding needed in the two width and height directions with least overhead possible
• emit warning if region is entirely out of bounds, i.e. read_region((-100, -100), 0, (50, 50))
• add a single call to np.pad just between here (Line 404), guarded only if needed: https://github.com/bayer-science-for-a-better-life/tiffslide/blob/56b5bbda36a2792a93b3eabd30c590f004835b41/tiffslide/tiffslide.py#L402-L405
• add tests for the specific edge-cases

This is a pretty well defined PR. Anyone up for tackling it? 😃

Cheers, Andreas

I am working on it.
A PR will open later.

[BenjaminIsaac0111]

Amazing, just saw this!

Thanks for the implementation! @ap-- @One-sixth

Wilson

[ap--]

Hey there,

Thanks for the bug report! ❤️
I'm a bit hesitant to implement this, but openslide definitely allows you to pad in all directions 🤔 ...

import numpy as np
import openslide

def _is_padding(im):
    return np.all(np.array(im) == 0)

slide = openslide.open_slide("some.svs")
slide.level_dimensions
>>> ((63743, 36055), (15935, 9013), (3983, 2253))

# oob positive
_is_padding(slide.read_region((1_000_000, 1_000_0000), 0, (10, 10)))
>>> True

# oob negative
_is_padding(slide.read_region((-1_000_000, -1_000_0000), 0, (10, 10)))
>>> True

# oob higher level than available
_is_padding(slide.read_region((0, 0), 1000, (10, 10)))
>>> True

# oob negative level
_is_padding(slide.read_region((0, 0), -1000, (10, 10)))
>>> True

# empty reads
slide.read_region((0, 0), 0, (0, 0))  # tile width and/or height can be 0
>>> <PIL.Image.Image image mode=RGBA size=0x0 at 0x7F926D19E220>

I think I'd like to default to not allow this (because I think it's a potential source of unwanted bugs). But we could offer a keyword argument and in addition some global setting to be able to allow the exact behavior above. Maybe even allowing to set the padding value.

This should also ship with more compatibility tests for openslide as mentioned in #1.

Would you like to contribute to implementing this functionality in a PR?
I could provide some feedback, or sketch out an initial draft.

Cheers,
Andreas 😃

[One-sixth]

@ap--
I had this problem recently too.
We need to keep the same padding method as openslide.
The coordinate system I use here uses negative numbers and functions that allow to exceed boundaries.

This makes it easy to slice the slide without running into annoying border issues.
In practice, we always do padding to keep the model input size.
Resize will causes abnormal image aspect ratio.

[ap--]

Hey everyone,

It does make sense to support this.

The implementation would be limited to these few lines:
https://github.com/bayer-science-for-a-better-life/tiffslide/blob/56b5bbda36a2792a93b3eabd30c590f004835b41/tiffslide/tiffslide.py#L358-L408

TODOs:

• add keyword only padding argument defaulting to True
• calculate padding needed in the two width and height directions with least overhead possible
• emit warning if region is entirely out of bounds, i.e. read_region((-100, -100), 0, (50, 50))
• add a single call to np.pad just between here (Line 404), guarded only if needed: https://github.com/bayer-science-for-a-better-life/tiffslide/blob/56b5bbda36a2792a93b3eabd30c590f004835b41/tiffslide/tiffslide.py#L402-L405
• add tests for the specific edge-cases

This is a pretty well defined PR. Anyone up for tackling it? 😃

Cheers,
Andreas

[One-sixth]

@ap--
How do you think about filling with any color? Let the parameter padding = (0, 0, 0) .

[ap--]

Let's start with padding with 0 only.
We'll think about user provided padding values in a separate issue.

Thank you so much for contributing!