Generalize SobelGradients to 3D and Any Axis

[bhashemian]

Fixes #5188

Description

This PR reimplement SobelGradients and SobelGradientsd using separable kernels and generalize it to images with any spatial dimension (2D, 3D, etc.) and option to calculate the gradient along any given axis.

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• New tests added to cover the changes.
• Integration tests passed locally by running ./runtests.sh -f -u --net --coverage.
• Quick tests passed locally by running ./runtests.sh --quick --unittests --disttests.
• In-line docstrings updated.

[KumoLiu]

Thanks for your update, it looks good to me now!

[wyli]

perhaps this need some refactoring for 3d as well

[KumoLiu]

Hi @drbeh,
I have just added unittest in CalcualteInstanceSegmentationMap which used SobelGradients and found some different effects between cv2 and SobelGradients.
The difference between cv2 and SobelGradients also cause CalcualteInstanceSegmentationMap can't get correct instance segmentation results. So I just used cv2 now. I will withdraw my approval of this PR and please check it.
Here is my test code.

import cv2
from monai.transforms.post.array import SobelGradients

image = np.zeros([20,20])
image[5:15, 5:15] = 1

filter = SobelGradients(kernel_size=5)
mo_ret = filter(image[None]).squeeze()
cv_ret_0 = cv2.Sobel(image, cv2.CV_64F, 0, 1, ksize=5)
cv_ret_1 = cv2.Sobel(image, cv2.CV_64F, 1, 0, ksize=5)
mo_ret = 1 - (mo_ret - np.min(mo_ret)) / (np.max(mo_ret) - np.min(mo_ret))
cv_ret_0 = 1 - (cv_ret_0 - np.min(cv_ret_0)) / (np.max(cv_ret_0) - np.min(cv_ret_0))
cv_ret_1 = 1 - (cv_ret_1 - np.min(cv_ret_1)) / (np.max(cv_ret_1) - np.min(cv_ret_1))

fig, ax = plt.subplots(1,5,figsize = (50,10))
ax[0].imshow(mo_ret[0], cmap='gray')
ax[1].imshow(mo_ret[1], cmap='gray')
ax[2].imshow(cv_ret_0, cmap='gray')
ax[3].imshow(cv_ret_1, cmap='gray')
ax[4].imshow(mo_ret[0] - cv_ret_0, cmap='gray')

Thanks!

as comment

[wyli]

I have just added unittest in CalcualteInstanceSegmentationMap which used SobelGradients and found some different effects between cv2 and SobelGradients

@drbeh and I were discussing this in another thread, opencv uses the separable implementation cv2.getDerivKernels(1, 0, 5), that's slightly different from what we have here. but we haven't looked into the actual segmentation outcome differences.

[KumoLiu]

Just FYI, I showed the instance segmentation result from CalcualteInstanceSegmentationMap. The results are calculated with all the same components except for Sobel gradient.
cv2 - result

MONAI - result

Thanks!

[JHancox]

Sorry - inadvertantly closed the PR

[bhashemian]

@wyli @Nic-Ma @KumoLiu any other comments here? Thanks

[wyli]

The api looks good to me

[Nic-Ma]

/black

[Nic-Ma]

/build

[wyli]

/build