Add median image filter as a transform

[dzenanz]

Is your feature request related to a problem? Please describe.
One problem is explained in #3178.

Describe the solution you'd like
Add a median filter based on @ebrahimebrahim's medianblur_with_gpu

Describe alternatives you've considered
#3178 mentions some alternatives, such as mean filtering. Those are suboptimal for many use cases.

Additional context
Model it after GaussianSmooth:

MONAI/monai/transforms/intensity/array.py

Lines 1139 to 1171 in ef68f16

	class GaussianSmooth(Transform):
	"""
	Apply Gaussian smooth to the input data based on specified `sigma` parameter.
	A default value `sigma=1.0` is provided for reference.
	Args:
	sigma: if a list of values, must match the count of spatial dimensions of input data,
	and apply every value in the list to 1 spatial dimension. if only 1 value provided,
	use it for all spatial dimensions.
	approx: discrete Gaussian kernel type, available options are "erf", "sampled", and "scalespace".
	see also :py:meth:`monai.networks.layers.GaussianFilter`.
	"""
	backend = [TransformBackends.TORCH]
	def __init__(self, sigma: Union[Sequence[float], float] = 1.0, approx: str = "erf") -> None:
	self.sigma = sigma
	self.approx = approx
	def __call__(self, img: NdarrayTensor) -> NdarrayTensor:
	img = convert_to_tensor(img, track_meta=get_track_meta())
	img_t, *_ = convert_data_type(img, torch.Tensor, dtype=torch.float)
	sigma: Union[Sequence[torch.Tensor], torch.Tensor]
	if isinstance(self.sigma, Sequence):
	sigma = [torch.as_tensor(s, device=img_t.device) for s in self.sigma]
	else:
	sigma = torch.as_tensor(self.sigma, device=img_t.device)
	gaussian_filter = GaussianFilter(img_t.ndim - 1, sigma, approx=self.approx)
	out_t: torch.Tensor = gaussian_filter(img_t.unsqueeze(0)).squeeze(0)
	out, *_ = convert_to_dst_type(out_t, dst=img, dtype=out_t.dtype)
	return out

[dzenanz]

Can someone assign this to me? I am starting work on it today.

[ebrahimebrahim]

Note that there may be a much better approach than the filter I wrote.

• Because mine has a giant sparse kernel (size $n^6$ when the filter size is $n$), it quickly becomes unwieldy as you increase the filter size. I wonder if there is a clever way around this. I sometimes see the MONAI folks implementing 2D or 3D convolutions in terms of 1D convolutions, but I never dug into how that actually works and when it can be done. Would be good to look at how some other convolution-based transforms in MONAI are implemented.
• My implementation is 3D and really we want something that is general to any spatial dimensionality. Actually I stole mine from the Kornia project; they had a 2D median filter that worked the same way.

[wyli]

I saw a similar implementation in tensorflow https://github.com/tensorflow/addons/blob/v0.17.0/tensorflow_addons/image/filters.py#L125-L200 given pytorch's convolution is up to 3d, perhaps we can support 1/2/3d as a starting point.

for the separable approximation, it seems it's useful in practice (for 2d) https://ieeexplore.ieee.org/document/4767047, maybe we can create another feature request for it after finishing this ticket, what do you think?