Issue Running Anisotropic Tutorial

[HanyingYan]

Hi,

Thank you for developing this tool! I’ve been trying to work around the indirect method by following your anisotropic tutorial. However, both segment_epithelial_organoid_2D_stacks.py and segment_single-cell_tracking_challenge_2D_stacks.py failed when running uSegment3D.aggregate_2D_to_3D_segmentation_indirect_method(), and I received the error message below.

Could you please let me know what might be causing this issue?

Thanks!

---

ValueError Traceback (most recent call last)
Cell In[35], line 2
1 # perform the indirect method. As we only have xy segmentations, we set the rest to empty []. We need to pass in the shape of the 'xy' view to get the dimensionality correct.
----> 2 segmentation3D, (probability3D, gradients3D) = uSegment3D.aggregate_2D_to_3D_segmentation_indirect_method(segmentations=[labels_xy,
3 [],
4 []],
5 img_xy_shape = img_preprocess.shape,
6 precomputed_binary=None,
7 params=indirect_aggregation_params,
8 savefolder=None,
9 basename=None)
11 uSegment3D_fio.save_segmentation(os.path.join(savefolder_seg,
12 'uSegment3D_epidermal_initial_labels.tif'), segmentation3D)

File ~/miniconda3/envs/segment3D/lib/python3.9/site-packages/segment3D/usegment3d.py:504, in aggregate_2D_to_3D_segmentation_indirect_method(segmentations, params, img_xy_shape, precomputed_binary, savefolder, basename)
502 if precomputed_binary is None:
503 if sigma > 0:
--> 504 mask_combine = usegment3D_filters.var_combine([ndimage.gaussian_filter((masks[0]>0)*1., sigma=sigma),
505 ndimage.gaussian_filter((masks[1]>0).transpose(1,0,2)*1., sigma=sigma),
506 ndimage.gaussian_filter((masks[2]>0).transpose(1,2,0)*1., sigma=sigma)],
507 ksize=params['combine_cell_probs']['ksize'], # awesome... high ksize is good
508 alpha = params['combine_cell_probs']['alpha'], # was 1.0
509 eps=params['combine_cell_probs']['eps'])
510 else:
511 mask_combine = usegment3D_filters.var_combine([(masks[0]>0)*1.,
512 (masks[1]>0).transpose(1,0,2)*1.,
513 (masks[2]>0).transpose(1,2,0)*1.],
514 ksize=params['combine_cell_probs']['ksize'], # awesome... high ksize is good
515 alpha = params['combine_cell_probs']['alpha'], # was 1.0
516 eps=params['combine_cell_probs']['eps'])

File ~/miniconda3/envs/segment3D/lib/python3.9/site-packages/segment3D/filters.py:704, in var_combine(arrays, ksize, alpha, eps)
701 import skimage.filters.rank as skfilters_rank #import entropy
702 import skimage.morphology as skmorph
--> 704 weights = np.array([1./(var_filter(array, ksize=ksize) + alpha) for array in arrays])
706 v = np.sum(weights*arrays, axis=0) / (weights.sum(axis=0)+eps)
708 return v

ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (3,) + inhomogeneous part.

[fyz11]

whats your operating system, python version and cellpose version (i.e. did you install cellpose-sam or another version) ? - I'll then try running those scripts on similar specs and get back to you. From the errors, it seems some sort of array size mismatch..... but i don't know why that would be the case here unless labels_xy.shape does not match img_preprocess.shape

[HanyingYan]

great, I am using HPC cluster (Linux) and the conda env

create –n segment3D python=3.9 
conda activate segment3D 
conda install -c conda-forge scikit-fmm 
pip install u-Segment3D  

with

# Name                     Version          Build            Channel
cellpose                   3.1.1.3          pypi_0           pypi
python                     3.9.25           h0dcde21_1

Thanks!

[fyz11]

wild. So, i wasn't able to reproduce the error on our HPC cluster (linux redhat) with those specifications. For me, the scripts ran fine...
I recall I did have experience with similar bug in cellpose3 in earlier versions of u-Segment3D.

Did you install u-Segment3D recently? Maybe try re-installing with pip install?

Were there potentially any modifications to the code?

The size of labels_xy and img_preprocess should be the same and is (50,512,512). - can you check thats the case? --- see screenshot.

[HanyingYan]

So the shape of img_preprocess is (50, 512, 5212) but the shape of labels_xy is (1,512, 512)

I believe I followed through the codes below before running labels_xy = all_masks_stack.copy()

Do you maybe know why? Thanks!

[fyz11]

somehow the loop ran once or there might be an overwrite.

just to be on the same page, could you share the exact script? - thanks.

[HanyingYan]

Sure, please find the notebook attached. Thanks!

segment_epithelial_organoid_2D_stacks.ipynb

[fyz11]

oh i see. When i look at your jupyter notebook. the code in cells 20,21,22 should all be in one cell.

The code should all be run under the for dd in np.arange(len(img_preprocess))[:]: loop. As written, you only perform segmentation for one xy slice, not the whole stack, therefore when you do the np.array() at the end, your shape is only (1,512,512) not (50,512,512) for labels_xy.

[HanyingYan]

I see. now it works! Thank you for the instruction!