Question: Why is default structuring element of 1-connectivity used by ndimage.label for getting the default markers(labeled local minima) for watershed segmentation?

[nbCloud91]

Description

In the implemented watershed.py segmentation, one can pass a connectivity argument, when local_minima is called by watershed (when markers=none) it passes the same connectivity value onto the local minima function, this generates the markers_bool, which then is used by ndimage.label() to generate the labeled markers. ndimage.label can take a structure argument and it defaults to 1-connectivity structure if none are given. IMO it is more intuitive to have ndimage.label also use the same connectivity which is passed by the user to the watershed function instead of always using the default 1-connectivity.
Below is the line in the source I am referring to:

scikit-image/skimage/segmentation/_watershed.py

Line 76 in 663e9e4

markers = ndi.label(markers_bool)[0]

Instead of the above would it not be better to have?

if connectivity==2:
   markers = ndi.label(markers_bool,
                       structure = [[1,1,1],
                                   [1,1,1],
                                   [1,1,1]])[0]

[jni]

@nbCloud91 you're right, we should definitely match the connectivity. Rather than an if statement, though, I would replace

markers = ndi.label(markers_bool)[0]

with

footprint = ndi.generate_binary_structure(markers_bool.ndim, connectivity)
markers = ndi.label(markers_bool, structure=footprint)[0]

[nbCloud91]

@jni could you make a PR for the above changes? I have not agreed to any CoC for this project and this seems like a small edit.

[jni]

@nbCloud91 you're right, it's a small fix, but we have lots of them! (See list of open issues/PRs 😫) If you have time to make a PR with the change + test, we would much appreciate it! If not, no worries, I definitely hope that we can get to it before the next release. 🤞

[divyank0]

I can raise this PR . :)

[divyank0]

@nbCloud91 I am trying to build a minimalistic test case here. It is not that simple.
to understand the situation, I checked for dimensions=2 and connectivity = 1 and 2.

But the output I am getting is same in previous version and new proposed version. If you can help me with a test case, I will raise the PR. I will also continue to build the test case

edit: The change is not failing any of the current tests. They are all passing.

Regards
Divyank

[nbCloud91]

Take for example this code:

import matplotlib.pyplot as plt
import skimage.morphology as sml
import skimage.segmentation as sgim
import numpy as np
from scipy import ndimage as ndi

# Generate a dummy BrightnessTemperature image
x, y = np.indices((406, 270))
x1, y1, x2, y2, x3, y3, x4, y4 = 200, 208, 300, 120, 100, 100, 340, 208
r1, r2, r3, r4 = 100, 50 ,40, 80
mask_circle1 = (x - x1)**2 + (y - y1)**2 < r1**2
mask_circle2 = (x - x2)**2 + (y - y2)**2 < r2**2
mask_circle3 = (x - x3)**2 + (y - y3)**2 < r3**2
mask_circle4 = (x - x4)**2 + (y - y4)**2 < r4**2
image = np.logical_or(mask_circle1, mask_circle2)
image = np.logical_or(image, mask_circle3)
image = np.logical_or(image, mask_circle4)

DummyBT = ndi.distance_transform_edt(image)
plt.imshow(DummyBT)

DummyBT_dis=np.around(DummyBT/12, decimals=0)*12
plt.imshow(DummyBT_dis)

Img_mask=np.where(DummyBT_dis==0,0,1)
plt.imshow(Img_mask)

plt.imshow(sgim.watershed(200-DummyBT_dis, mask=Img_mask, connectivity=2, compactness=0.01))

You can see the issue with the last image, the small stray triangular segmentation. With two connectivity, this artifact should not appear.