This repository contains the scripts needed to reproduce the eosinophil cell counting analysis. The code requires assumes that the folder structure is set up as follows:

• GICell
  • code
  • images
  • output
  • points
• GIOrdinal
  • data

There are three main shell scripts which need to be run in a specific order (shown below). A description of each file in the script is summarized below. Please review the shell scripts themselves as they provide additional commentary on the python scripts.

1. 0_pipeline_hp.sh: Processes data and trains models across different hyperparameter configurations. Wait for 5_pipeline_nohup.sh to finish running before moving onto the next step.
   1. set_env.sh: This will look to see if your conda installation has an environment called QuPath. If the environment is not found, it will be created with the necessary packages found in conda_env.txt. Note that if you do not have miniconda installed, you may need to change line 4.
   2. 1a_check_crop.py: This is a sanity check script to make sure that there are labels found for the different crops. Change the find_dir_cell function to the path the directory structure shown above.
   3. 1b_check_cuda.py: Makes sure that CUDA is configured properly and that PyTorch can connect to a GPU device.
   4. 1c_check_anno.py: Provides a print read-out of the different annotations and which folders they live in (as well if there are any images that do not yet have annotations).
   5. 1d_check_trans.py: Makes sure that the image encoder can apply loss-less rotations and flips (up to some floating point errors).
   6. 2_process_xy.py: Loads all annotations and crops and saves an pickled dictionary of data for each annotator.
   7. 3_data_split.py: Assigns images into training, validation, and test.
   8. 4_run_mdl.py: Checks that model class works with --check_model argument (will run for one epoch and produce no output)
   9. 5_pipeline_nohup.sh: Runs pipeline_unet_hp.sh with no hang up, which will call 4_run_mdl.py with different hyperparameter configurations for learning rate, batch size, and model complexity. The dir_code, path_conda, and dir_nohup variables will need to be changed for a new machine.
2. 6_pipeline_inf.sh: Determines which hyperparameters are optimal and retrains models with these settings. Wait for 8_pipeline_best.sh to finish running before moving onto the next step.
   1. 7_explore_hp.py: Determines what are the "best" hyperparameters in terms of AUROC
   2. 8_pipeline_best.sh: Calls in pipeline_unet_best.sh with no hang up, and uses the output from the previous script to set hyperparameters for exact training schedule.
3. 9_analyze_test.sh: Uses "best" models to carry out inference on test sets.
   1. 10_rename_mdls.py: Changes the hashed model name to best_eosin.pickle.
   2. 11_explore_test.py: Gets aggregate performance as well as individualized inferences on different datasets using the best model.
   3. 12_explore_inter.py: Compares inter-annotator variability of pathologists, to that of the model and pathologists.