measure.regionprops_table with properties=['centroid_weighted'] fails on images with >1 color channel.

[aslyon]

Description:

Calling skimage.measure.regionprops_table with properties=['centroid_weighted'] on an image with >1 color channel raises ValueError: setting an array element with a sequence. I expected to get a dictionary with keys centroid_weighted-n-c where n are the indices of the spatial dimensions and c are the indices of the color channels.

Way to reproduce:

import numpy as np
import skimage

# Create 3D label image with a sphere
label_img = np.zeros((9, 9, 9), dtype='uint8')
label_img[:, :, :] = skimage.morphology.ball(radius=4)
n_pixels = label_img.sum()

# Create random number generator
rng = np.random.default_rng(seed=123)

# Create intensity image of sphere filled with random values
intensity_img = np.zeros((9, 9, 9, 3))
intensity_img[:, :, :, 0][label_img == 1] = rng.uniform(low=10, high=20, size=n_pixels)
intensity_img[:, :, :, 1][label_img == 1] = rng.uniform(low=10, high=20, size=n_pixels)
intensity_img[:, :, :, 2][label_img == 1] = rng.uniform(low=10, high=20, size=n_pixels)

# Measure weighted centroid for 1 color channel - this works
skimage.measure.regionprops_table(label_img, intensity_img[:, :, :, 0:1], properties=['centroid_weighted'])
# output: {'centroid_weighted-0': array([4.0115177]), 'centroid_weighted-1': array([3.98670859]), 'centroid_weighted-2': array([3.99605087])}

# Measure weighted centroid for 2 color channels - this DOESN'T work
skimage.measure.regionprops_table(label_img, intensity_img[:, :, :, 0:2], properties=['centroid_weighted'])

# Measure weighted centroid for 3 color channels - this DOESN'T work
skimage.measure.regionprops_table(label_img, intensity_img[:, :, :, 0:3], properties=['centroid_weighted'])

The ValueError with traceback is below. The error is the same whether 2 or 3 color intensity image is used.

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
TypeError: only size-1 arrays can be converted to Python scalars

The above exception was the direct cause of the following exception:

ValueError                                Traceback (most recent call last)
Cell In[86], line 26
     16 intensity_img[:, :, :, 2][label_img == 1] = rng.uniform(low=10, high=20, size=n_pixels)
     18 # Measure weighted centroid for 1 color channel - this works
     19 # skimage.measure.regionprops_table(label_img, intensity_img[:, :, :, 0:1], properties=['centroid_weighted'])
     20 # output: {'centroid_weighted-0': array([4.0115177]), 'centroid_weighted-1': array([3.98670859]), 'centroid_weighted-2': array([3.99605087])}
   (...)
     24 
     25 # Measure weighted centroid for 3 color channels - this DOESN'T work
---> 26 skimage.measure.regionprops_table(label_img, intensity_img[:, :, :, 0:3], properties=['centroid_weighted'])

File /Applications/miniconda3/envs/idpimg/lib/python3.10/site-packages/skimage/measure/_regionprops.py:1045, in regionprops_table(label_image, intensity_image, properties, cache, separator, extra_properties, spacing)
   1041     out_d = _props_to_dict(regions, properties=properties,
   1042                            separator=separator)
   1043     return {k: v[:0] for k, v in out_d.items()}
-> 1045 return _props_to_dict(
   1046     regions, properties=properties, separator=separator
   1047 )

File /Applications/miniconda3/envs/idpimg/lib/python3.10/site-packages/skimage/measure/_regionprops.py:877, in _props_to_dict(regions, properties, separator)
    875     rp = regions[k][prop]
    876     for i, loc in enumerate(locs):
--> 877         column_data[k, i] = rp[loc]
    879 # add the columns to the output dictionary
    880 for i, modified_prop in enumerate(modified_props):

ValueError: setting an array element with a sequence.

Version information:

3.10.8 (main, Nov 11 2022, 08:11:25) [Clang 12.0.0 ]
macOS-10.16-x86_64-i386-64bit
scikit-image version: 0.20.0
numpy version: 1.22.3

[jni]

Thanks for the report, @aslyon! The issue is here:

scikit-image/skimage/measure/_regionprops.py

Lines 857 to 860 in c7fb481

	if isinstance(rp, np.ndarray):
	shape = rp.shape
	else:
	shape = (len(rp),)

It turns out that the output of weighted_moments is an array-like, a tuple of arrays, rather than an array, so instead of inferring a 2D array here, we infer a linear set of numbers. This just happens to work when the arrays contain one element, but fails when there's more, as you have observed. Here's how I modified your example to show these outputs:

import numpy as np
import skimage

# Create 3D label image with a sphere
label_img = np.zeros((9, 9, 9), dtype='uint8')
label_img[:, :, :] = skimage.morphology.ball(radius=4)
n_pixels = label_img.sum()

# Create random number generator
rng = np.random.default_rng(seed=123)

# Create intensity image of sphere filled with random values
intensity_img = np.zeros((9, 9, 9, 3))
intensity_img[:, :, :, 0][label_img == 1] = rng.uniform(low=10, high=20, size=n_pixels)
intensity_img[:, :, :, 1][label_img == 1] = rng.uniform(low=10, high=20, size=n_pixels)
intensity_img[:, :, :, 2][label_img == 1] = rng.uniform(low=10, high=20, size=n_pixels)

# Measure props for no channel axis:
props0 = skimage.measure.regionprops(label_img, intensity_img[..., 0])[0]
print(f'no channel:\n {props0.centroid_weighted}')

# Measure props for single channel
props1 = skimage.measure.regionprops(label_img, intensity_img[..., 0:1])[0]
print(f'one channel:\n {props1.centroid_weighted}')

# Measure props for multiple channels
propsm = skimage.measure.regionprops(label_img, intensity_img)[0]
print(f'multichannel:\n {propsm.centroid_weighted}')

which gives:

no channel:
 (4.011517704412646, 3.9867085865741485, 3.996050865015718)
one channel:
 (array([4.0115177]), array([3.98670859]), array([3.99605087]))
multichannel:
 (array([4.0115177 , 4.03724327, 3.98084178]), array([3.98670859, 4.01261718, 3.96234087]), array([3.99605087, 4.0387654 , 3.99636875]))

So, it looks like we need a better way of identifying the type and shape of the output of each property when turning it into a table. My first instinct is to convert any non-scalar, non-object property to an array. I might give that a go.

[jni]

Attempted fix in #6861. Please check locally with your data @aslyon!

[aslyon]

Working correctly now, thanks so much for the rapid response @jni !