Refactor morphological functions as transform (or transform utils)

[mingxin-zheng]

Describe the bug

The original discussion was in PR comment:
#7893 (comment)

Will a refactor of the erodeand dilate functions make it easier in different applications/networks/pre-processing?

https://github.com/Project-MONAI/MONAI/blob/dev/monai/apps/generation/maisi/utils/morphological_ops.py

Expected behavior

class Erode3D(Transform):
   ...

# erode with a 3D filter when needed
input_img = torch.rand(b, l, h, d)
eroder = Erode3D(...)
output = eroder(input_img) 

[mingxin-zheng]

Since this operation is used by MAISI and VISTA3D both, do you think it will help? @heyufan1995 @Can-Zhao @guopengf ?

[KumoLiu]

Also cc @yiheng-wang-nv @ericspod

[ericspod]

We should look at what other requirements there are but I think it would make sense to have a morphological operations transform to encapsulate them.

[aylward]

+1 for making them transforms.

Should MONAI use ITK's / SimpleITK's implementations? They offer a multitude of options: binary, grayscale, open, close, generalized - and all are multi-threaded, I believe some contain logic to specifically to only perform convolution only at an object's edges to significantly speed results compared to full convolution, and other developers have created GPU/CUDA accelerated versions but they aren't in ITK by default.

[ericspod]

We should look into what the implementation would look like, it's definitely possible to at least support using ITK implementations, though the virtue of what was added with MAISI was that it is only Pytorch. We could support using ITK if present but then default to Pytorch implementations if not.

[qic999]

I meet the error ModuleNotFoundError: No module named 'monai.transforms.morphological_ops'

[KumoLiu]

I meet the error ModuleNotFoundError: No module named 'monai.transforms.morphological_ops'

Hi @qic999, it should be monai.transforms.utils_morphological_ops.

[qic999]

I meet the error ModuleNotFoundError: No module named 'monai.transforms.morphological_ops'

Hi @qic999, it should be monai.transforms.utils_morphological_ops.

OK. Thanks. This error occurs on line 24 of https://github.com/Project-MONAI/tutorials/tree/main/generation/maisi/scripts/utils.py.

[KumoLiu]

OK. Thanks. This error occurs on line 24 of https://github.com/Project-MONAI/tutorials/tree/main/generation/maisi/scripts/utils.py.

Thanks, will fix it in the following prs.

[qic999]

OK. Thanks. This error occurs on line 24 of https://github.com/Project-MONAI/tutorials/tree/main/generation/maisi/scripts/utils.py.

Thanks, will fix it in the following prs.

new error: ModuleNotFoundError: No module named 'monai.transforms.utils_morphological_ops'
my env is

[mingxin-zheng]

Hi @aylward , the current implementations of erode/dilate are basically padding and conv2d/3d with torch

MONAI/monai/transforms/utils_morphological_ops.py

Line 90 in 4877767

def get_morphological_filter_result_t(mask_t: Tensor, filter_size: int | Sequence[int], pad_value: float) -> Tensor:

Currently the usage of erode/dilate needs to be on GPUs, and it will be more work to do if we deviate from the torch implementation.

In my opinion, the role of itk is more like another backend on top of numpy and torch. So my gut feeling of ITK integration is that a wrapper that can bridge torch and itk/simple-itk, but I could be wrong here.

[mingxin-zheng]

OK. Thanks. This error occurs on line 24 of https://github.com/Project-MONAI/tutorials/tree/main/generation/maisi/scripts/utils.py.

Thanks, will fix it in the following prs.

new error: ModuleNotFoundError: No module named 'monai.transforms.utils_morphological_ops' my env is

Hi @qic999 , you'll need the dev version or the weekly build. The morphological operations are added recently and they are not available in monai 1.3.2, which is the version you are using. Check this page for installation guide:

https://docs.monai.io/en/stable/installation.html

If my comment above can't resolve your issue, please open a new issue so that the team can better support you. I would like to keep this thread simply focus the implementation plan. Thank you!