3D Feature Extraction debugging

[nrepina]

This PR addresses a few bug fixes in the Feature Extraction task:

• Bug fix in feature extraction to handle child object masking by parent. Previously masking was performed by selecting the label of child object that corresponds to parent object. This approach works for masking of objects of the same child/parent class (e.g. organoids masked by organoids), but results in an empty label image for child/parent relations, except in the rare case that the parent object happens to have the same label as one of the children objects. Proposed fix is to load the parent object segmentation image and mask by the binary image. This solution handles masking for both child/parent relations and objects of the same class, however to reduce code modification and backwards compatibility the previous method was kept in the case that a masking label name is not supplied. Current tests pass, but should be extended to check for the child/parent masking case. @jluethi See TODO notes in task for better handling of edge cases (e.g. upscaling in case of different array shapes)
• Improved logging and check for empty arrays.
• Fix Anndata dtype deprecation warning.
• Cast current region label to float prior to int conversion to handle ValueError in case of ROI label that is stored as string of float.

@jluethi Please have a look and we can decide on the best approach for these fixes before merging into main.

[codecov-commenter]

:warning: Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

Attention: Patch coverage is 0% with 15 lines in your changes missing coverage. Please review.

Project coverage is 0.26%. Comparing base (6ac2465) to head (4c8ebe3). Report is 11 commits behind head on main.

Files	Patch %	Lines
...tiplex/fractal/scmultiplex_feature_measurements.py	0.00%	15 Missing :warning:

:exclamation: Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

```diff @@ Coverage Diff @@ ## main #122 +/- ## ======================================== - Coverage 0.27% 0.26% -0.01% ======================================== Files 54 55 +1 Lines 4805 4942 +137 ======================================== Hits 13 13 - Misses 4792 4929 +137 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[jluethi]

Hey @nrepina

This looks good, great to hear that it's working now. I looked through the tests and we really only test the one masking case where we measure on the same objects that are also our mask. So far, all measurement tests just use the original, very tiny dataset. And that test dataset only has one set of labels (only nuclei, no organoids).

We now also have the linking dataset available in the test suite that you've added @nrepina (see https://zenodo.org/records/10683087). I'll work towards getting a test on that one running, because it actually contains nuclei & organoids => we can actually test for this case with the new test dataset :)

The upsampling already seems nicely integrated, though testing this further should wait for ngio flexibility.

Thus, I think the TODO list is:

• [ ] Create a test case that fails with old version (or with wrong selected table), but passes in new version
  • [ ] Use new https://zenodo.org/records/10683087 test dataset (available as linking_zenodo_zarrs)
• [ ] Get selection of label image to load from table attrs
• [ ] Validate that table is masking ROI table

[jluethi]

Turns out, using the https://zenodo.org/records/10683087 dataset for this tests won't be trivial atm (until there is a better fix for https://github.com/fractal-analytics-platform/fractal-tasks-core/issues/771). The test data contains organoids that have a lower resolution label image than the nuc.

I'll give it a quick shot to generate a "psuedo-org" label image on-the-fly to be able to run those tests

[nrepina]

Hey @jluethi, great, sounds like a plan. It has been on my to-do list to update the Zenodo test dataset with an object segmentation that matches the child segmentation level. I will get to that in the coming days and this should resolve the testing case until Fractal has a better solution for upscaling :) Will ping you once it is updated.

[nrepina]

@jluethi update of Zenodo dataset is finished! Now we should be able to cover the use case of masked single-cell feature extraction with a test dataset. I updated the scmpx integration tests, it's now ready for you to update the feature extraction tests and merge to main :)