embryo_binary_segmentation

CNN-based model for 3D segmentation of mouse embryo

---

Installation

You can install embryo_binary_segmentation via pip:

pip install embryo_binary_segmentation

To install latest development version :

pip install git+https://github.com/GuignardLab/embryo_binary_segmentation.git

Data structure

compliance with the data structure is necessary for correct data loading

It is important that the folder with images has FUSE in its name, and the folder with masks has SEG in its name.

Each folder must have the following structure:

• Train
  • JLM_12
    • FUSE
      • e1.tif
      • e2.tif
      • ...
    • SEG
      • e1.tif
      • e2.tif
      • ...
  • Woon_7
    • FUSE
      • e1.tif
      • e2.tif
      • ...
    • SEG
      • e1.tif
      • e2.tif
      • ...

The alphabetical order of image files and masks must match.

Parameters

DATA_PARAMS = {

"data_path": folder with the Train and Val datasets,

"binarize": if you need to binarize masks before training,

"target_size": size of cropped images,

"patch_size": size of patches (less size, easy trainig),

"augmentations": if you want to apply augmentations

}

FINE_TUNING = {

"upload_model_path": path to saved model,

"old_steps": if you want to continue epochs count,

}

TRAINING_PARAMS = {

"loss": loss function,

"learning_rate": 1e-4,

"batch_size": 8,

"epochs": total amount,

"save_model_path": folder where you will save best model,

"fine_tuning":True,

"save_each": if you want to save each 5th model weights,

}

TEST_PARAMS = {

"data_path":"folder with test data",

"binarize":False,

"target_size":[64, 512, 512],

"patch_size":[32, 512, 512],

"batch_size": 2,

"load_model_path": path to the model weights,

"load_csv_path": path to the csv file with losses by epochs

}

PRED_PARAMS = {

"data_path": folder with data for final prediction,

"final_load_model_path": path to the model weights,

"batch_size": don't use big number because you predict on the full size images,

"save_pred_path": where you want to save predictions

}

Contributing

Contributions are very welcome. Tests can be run with tox, please ensure the coverage at least stays the same before you submit a pull request.

License

Distributed under the terms of the MIT license, "embryo_binary_segmentation" is free and open source software

Issues

If you encounter any problems, please file an issue along with a detailed description.

---

This library was generated using Cookiecutter and a custom made template based on @napari's cookiecutter-napari-plugin template.