I just complete an early version of a simple ML-based 2-D nucleus/cell segmentation tool (as jupyter notebook).
It is based on:
a) The U-Net from Cellprofiler
b) The StarDist Model
c) The Cellpose Model
Typically, the user provides a .lif file, or a folder containing many .tif files as input (e.g. Z-Stacks).
In the simplest workflow you will just need to provide the filepath.
In this case the notebook will:
a) Run six standard pre-processing methods (1…6)
b) Run on these intermediate results the segmentation using (a,b,c)
c) Post-process images (filtering for min and max cell size).
Results are saved as segmentation masks (first folder), and overlays displaying the cell border as red line (second folder).
Typically, you would run a test on a subset of your images and simply look at the results (which you can open as Z-stack in Fiji).
The settings used to create the best result, might be used to segment all your images.
If none looks good, you might train you own model (elsewhere), and make it then available through this tool.
If you want to test the tool now, you should have a GPU and know how to create a Python environment from an .yml file.
At current creating tiles is not yet implemented.
The largest tested file size is 2048 x 2048 x n.
Please let me know if you have any further questions.
I would appreciate a lot, if some of you were interested to beta-test my tool.
Please message me & I will invite you to the repository. I would then kindly ask you to give me feedback & and share some test data. The tool might be further developed in many directions. Although I might not be able to follow up on most of them, I would encourage you to contribute to the tool whatever you would need for your application.
Thanks a lot &