W-Net, U-Net and SegNet for Corneal Endothelium Image Segmentation

Repository contains implementation of convolutional neural network models: W-Net, U-Net and SegNet based for corneal endothelium image segmentation.

Terms of use

The source code and image dataset may be used for non-commercial research provided you acknowledge the source by citing the following paper:

Adrian Kucharski, Anna Fabijańska, CNN-watershed: A watershed transform with predicted markers for corneal endothelium image segmentation, Biomedical Signal Processing and Control, Volume 68, 2021, 102805, ISSN 1746-8094, https://doi.org/10.1016/j.bspc.2021.102805

@article{Kucharski2021,
  doi = {10.1016/j.bspc.2021.102805},
  url = {https://doi.org/10.1016/j.bspc.2021.102805},
  year = {2021},
  month = jul,
  publisher = {Elsevier {BV}},
  volume = {68},
  pages = {102805},
  author = {Adrian Kucharski and Anna Fabija{\'{n}}ska},
  title = {{CNN}-watershed: A watershed transform with predicted markers for corneal endothelium image segmentation},
  journal = {Biomedical Signal Processing and Control}
}

Prerequisites

Code was tested on Windows 10 64-bit with Python 3.8, and TensorFlow 2.3.1.

Folder tree compatible with default config.ini file.


main_dir
├── Postprocess
│   └── postprocess.py
├── Predicted_images        # contains predicted images, run predict.py
│   └── Fold_0
│   │   └── UNet
│   │   └── SegNet
│   │   └── WNet
│   └── Fold_1
│   │   └── ...
│   └── Fold_2
│   │   └── ...
│   └── Fold_3
│   │   └── ...
│   └── Fold_4  
│       └── ...
├── Trained_model           # contains trained model and training history
│   └── UNet                
│   └── SegNet                
│   └── WNet              
├── Training_data
│   ├── markers             # markers generated from images from ./gt_all/ images
│   ├── gt_all              # ground truth images from http://bioimlab.dei.unipd.it/Endo%20Aliza%20Data%20Set.htm
│   ├── gt                  # put ground truth images here
│   ├── org                 # put original images here
│   └── field               # put region of interest images here
├── config.ini
├── history_show.py
├── models.py
├── others.py
├── predict.py
├── prepare_dataset.py
├── readme.md
└── training.py

How to use

Framework is designed to train and predict images with cross_validation setup. Default numbers of folds is 5 (80% train, 20% predict).

Config the config.ini file
Run prepare_dataset.py
Run training.py
Run predict.py

Content