Cell Segmentation ML Scripts

Description

This repo is still under work and as such is rapidly changing and a little messy for the sake of development time. The main goal is to currently get the segmentations masks we need and then I will spend time cleaning up and optimizing scripts for speed and reliability. If you would like to use any of these scipts and they are not working submit an issue and I can help as needed.

Scripts useful for training Cellpose and Omnipose models and running predictions on using trained models. These are used for 3D microscopy images in TIF format, however can be adapted to be used on other structured data.

Getting Started

Dependencies

• Python 3.8 or higher
• Cellpose and/or Omnipose, please follow their instructions for installing each.
• Additionally Required Libraries: Ray, Scikit-Optimize, Simple-ITK, Dask/Dask-Cuda
• OS: Windows/Linux/MacOS

Executing program

preprocessing.py

• Description: Preprocesses 3D TIF image data for machine learning tasks, supporting cropping and dataset splitting.
• Arguments:
  • --dir (str, absolute path): The directory containing input TIFF images.
  • --test_size (float): Proportion of the dataset to use as the test set.
  • --crop_size (list of int, separated by spaces): The crop size dimensions [z, y, x].
  • --strides (list of int, separated by spaces, optional): The strides for cropping in each dimension.
  • --save_name (str): Base name for output directories.
  • --remove_label (int): A label in the mask to be removed.
• Usage:

  python /absolute/path/to/preprocessing.py --dir /path/to/directory --test_size 0.2 --crop_size 1 64 64 --strides 1 32 32 --save_name processed --remove_label 0
  

postprocessing.py

• Description: Does nothing currently.
• Arguments:
  • --dir (str, absolute path): The directory containing input TIFF images.
• Usage:

  python /absolute/path/to/postprocessing.py --dir /path/to/directory
  

omnipose_run.py

• Description: Runs image segmentation using the OmniPose model.
• Arguments:
  • --image (str, absolute path): Path to the input image.
  • --mask (str, absolute path, optional): Path to the input mask.
  • --model (str, absolute path): Path to the OmniPose model.
  • --save_name (str): Name for the output.
• Usage:

  python /absolute/path/to/omnipose_run.py --image /path/to/image --mask /path/to/mask --model /path/to/model --save_name output
  

cellpose_run.py

• Description: Applies the CellPose model for image segmentation.
• Arguments:
  • --dir (str, absolute path): The directory for input/output.
  • --image_path (str, absolute path): Path to the input image.
  • --model (str, absolute path): Path to the CellPose model.
  • --pretrained (str, absolute path, optional): Path to a pretrained model.
  • --chunks (list of int, separated by spaces, optional): Chunk sizes for processing.
  • --kwargs (str, JSON format): Additional arguments for the CellPose model.
  • --save_name (str): Name for the output.
  • --batch_num (str, optional): Batch number.
  • --split (str, optional): Data split method.
• Usage:

  python /absolute/path/to/cellpose_run.py --dir /path/to/directory --image_path /path/to/image --model /path/to/model --pretrained /path/to/pretrained --chunks 500 500 --kwargs '{"diameter": 30, "do_3D": true}' --save_name output --batch_num 1 --split method
  

clean_training_data.py

• Description: Cleans training data using Geodesic Active Contour to refine cell mask labels, with options for background removal and visualization. Particularly useful for labels that have been upscaled and to possibly address over segmentation or small labels to some extent.
• Arguments:
  • --image_path (str, absolute path): Path to the input image.
  • --mask_path (str, absolute path): Path to the input mask.
  • --num_chunks (int, optional): Number of chunks for parallel processing.
  • --object_store_memory (int, optional): Memory for Ray's object store.
  • --background (int, optional): Specify to remove background if labeled.
  • --visualize (flag, optional): Enable visualization.
• Usage:

  python /absolute/path/to/clean_training_data.py --image_path /path/to/image --mask_path /path/to/mask --num_chunks 4 --object_store_memory 5000000000 --background 1 --visualize
  

batch_script_generator.py

• Description: Generates and executes batch scripts for testing multiple instances of Cellpose predictions using different settings. Will probably update Cellpose script to use similar optimization techniques used in omnipose_run.py to test multiple settings.
• Arguments:
  • --image_path (str, absolute path): Path to the input image.
  • --save_name (str): Name for the output.
• Usage:

  python /absolute/path/to/batch_script_generator.py --image_path /path/to/image --save_name output
  

omnipose_loss_visualization.ipynb

• Description: Jupyter notebook for visualizing the loss of the OmniPose model.
• How to run: Open the notebook in Jupyter and run the cells.