Use stardist segmentation for ellipsoid fitting

[stefanhahmann]

No description provided.

[stefanhahmann]

User story

The current implementation of ellipsoid fitting is rather slow (~5 hours for 500.000 spots) and also imprecise. Both is unsatisfiable for a user who wants to run it. Thus, it would be nice, if the ellipsoid fitting could exploit the GPU, if an NVIDIA graphics card is available. I would be desirable to achieve higher performance and more precise results.

Acceptance criteria

• Additional Ellipsoid fitting option that
  • uses JDLL: https://github.com/bioimage-io/JDLL
  • cuda: https://developer.nvidia.com/cuda-toolkit
  • cuDNN: https://developer.nvidia.com/cudnn
• Runs significantly faster than current ellipsoid fitting when performed on GPU
• Exploits stardist segmentation algorithms - https://github.com/stardist/stardist

Tasks

• Prototypical testing by https://github.com/Noam-Dori
  • Load resulting code into project - cf.: stardist_test.zip
• Install GPU drivers
  • Install cuda (https://developer.nvidia.com/cuda-toolkit)
  • Install cuDNN (https://developer.nvidia.com/cudnn)
• One time procedure: download tensorflow engine via the outcommented code
• Adapth paths to file system
• Test, if the code runs stabily (multiple times). On the machine of https://github.com/Noam-Dori the segmentation crashed sometimes
• Partion/crop images before handing them over to prediction. A GPU memory error occurs, if one complete timepoint is processed by stardist at the same time. Partioning and cropping may solve this.
• Filter StarDist results. In the StarDist 2D ImageJ plugin, there is some post-processing of StarDist results (i.e. non-maximum suppression), which does not exist in the 3D translation and results in too many segments.
• Integrate loop for multiple timepoints
• Handle the disk leak C:\User\AppData\Local\Temp\tensorflow_* -> delete those files after running (it is a downloaded library from online used in processing)
• Add method to derive variance/covariance params from set of Pixels, cf.: Compute ellipsoid for segmentation result #9
• Integrate that into the plugin

Additional information

• StarDist publicly available training+test data: https://github.com/stardist/stardist/releases/download/0.3.0/demo3D.zip

  • Timelapse with 27 timepoints as HDF5 dataset:
    stardist.zip

• StarDist3D model to be used with JDLL: adapted_stardist_model.zip

  • Model has been translated from pre-trained StarDist Model (contained here: https://github.com/stardist/stardist/releases/download/0.3.0/demo3D.zip) using this Python script: convert_stardist_h5_to_pb.zip