bioimage-io / collection-bioimage-io

(deprecated in favor of bioimage-io/collection) RDF collection for BioImage.IO

5 stars 9 forks source link

StarDist Fluorescence Nuclei Segmentation #596

Open utterances-bot opened 1 year ago

utterances-bot commented 1 year ago

StarDist Fluorescence Nuclei Segmentation

Bioimage.io -- an AI model repository for deep learning.

https://bioimage.io/?partner=ilastik&type=model&id=10.5281%2Fzenodo.6348084

manerotoni commented 1 year ago

hello, I am testing the stardist model in ilastik. I am not so sure what is the results you should get. The probabilities only, the label masks. It gives a prediction of 33 labels (why 33?, there are more of it) but none of those corresponds to the label mask. Label 1 could be the probability but it has a checkboard pattern. (Sorry I can't upload an image here)

manerotoni commented 1 year ago

Ilastik version 1.4.0, GPU Model that do not do instance segmentation seems to work well.

manerotoni commented 1 year ago

The 10.5281/zenodo.6348084 output looks as the sample outputs. I guess needs to do a thresholding of the probabilities and CCA. For this particular model I am little puzzled in how to do it with the checkboard patterns in label 1 and the other labels. I think the other labels may be the predicted starconvex angles?

FynnBe commented 1 year ago

Hello @manerotoni , yes, the stardist model output is very specific to the stardist method. The stardist postprocessing is (not yet) available in ilastik. The full stardist pipeline can already be executed

in Python: https://github.com/stardist/stardist
in DeepImageJ (Fiji Plugin): https://www.nature.com/articles/s41592-021-01262-9

You can also use the Python stardist package to perfrom the postprocessing only and keep using ilastik for the model inference.