This is the official repository of the paper "RRWNet: Recursive Refinement Network for Effective Retinal Artery/Vein Segmentation and Classification", by José Morano, Guilherme Aresta, and Hrvoje Bogunović, published in Expert Systems with Applications (2024).

[arXiv] ESWA] [BibTeX]

• Basic inference code.
• Training code.
• Preprocessed training and testing data.
• Data preprocessing code.
• Evaluation code.
• Better documentation.
• Easier execution of the evaluation.

The predictions for the different datasets as well as the weights for the proposed RRWNet model can be found at the following link:

• https://drive.google.com/drive/folders/1Pz0z-OxzEft5EWGbZ3MeeqQNWqvv2ese?usp=sharing

The model trained on the RITE dataset was trained using the original image resolution, while the model trained on HRF was trained using images resized to a width of 1024 pixels. The weights for the RITE dataset are named rrwnet_RITE_1.pth, while the weights for the HRF dataset are named rrwnet_HRF_0.pth. Please note that the size of the images used for training is important when using the weights for predictions.

The code was tested using Python 3.10.10. The following instructions are for setting up the environment using pyenv and pip. However, the code should work with other Python versions and package managers. Just make sure to install the required packages listed in requirements.txt.

Install pyenv.

curl https://pyenv.run | bash

Install clang. E.g.:

sudo dnf install clang

Install Python version 3.10.10.

CC=clang pyenv install -v 3.10.10

Create and activate Python environment.

~/.pyenv/versions/3.10.10/bin/python3 -m venv venv/
source venv/bin/activate  # bash
. venv/bin/activate.fish  # fish

Update pip.

pip install --upgrade pip

Install requirements using requirements.txt.

pip3 install -r requirements.txt

You can preprocess the images offline using the preprocessing.py script. The script will enhance the images and masks and save them in the specified directory.

python3 preprocessing.py --images-path data/images/ --masks-path data/masks/ --save-path data/enhanced

To get predictions using the provided weights, run the get_predictions.py script. The script will save the predictions in the specified directory. If the images were not previously preprocessed, you can use the --preprocess flag to preprocess the images on the fly.

python3 get_predictions.py --weights rrwnet_RITE_1.pth --images-path data/images/ --masks-path data/masks/ --save-path predictions/ --preprocess

You can refine existing predictions (e.g., from a different model) using the same get_predictions.py script. The script will save the refined predictions in the specified directory. Just make sure to provide the path to the predictions and the weights to be used for the refinement. Also, do not forget to use the --refine flag and do not use the --preprocess flag.

python3 get_predictions.py --weights rrwnet_RITE_refinement.pth --images-path data/U-Net_predictions/ --masks-path data/masks/ --save-path refined_predictions/ --refine

All training code can be found in the train/ directory. The training script is train.py, and the configuration file, with all the hyperparameters and command line arguments, is config.py. Please follow the instructions in train/README.md to train the model.

If you have any questions or problems with the code or the paper, please do not hesitate to open an issue in this repository or contact me at [email protected].

If you use this code, the weights, the preprocessed data, or the predictions in your research, we would greatly appreciate it if you cite our work:

@article{morano2024rrwnet,
    title={RRWNet: Recursive Refinement Network for Effective Retinal Artery/Vein Segmentation and Classification},
    author={Morano, Jos{\'e} and Aresta, Guilherme and Bogunovi{\'c}, Hrvoje},
    journal={Expert Systems with Applications},
    year={2024},
    doi={10.1016/j.eswa.2024.124970}
}

Also, if you use any of the public datasets used in this work, please cite the corresponding papers:

• RITE
  • Images: Staal, Joes, et al. "Ridge-based vessel segmentation in color images of the retina." IEEE transactions on medical imaging 23.4 (2004): 501-509.
  • Annotations: Hu, Qiao, Michael D. Abràmoff, and Mona K. Garvin. "Automated separation of binary overlapping trees in low-contrast color retinal images." Medical Image Computing and Computer-Assisted Intervention–MICCAI 2013: 16th International Conference, Nagoya, Japan, September 22-26, 2013, Proceedings, Part II 16. Springer Berlin Heidelberg, 2013.
• LES-AV
  • Images and annotations: Orlando, José Ignacio, et al. "Towards a glaucoma risk index based on simulated hemodynamics from fundus images." Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11. Springer International Publishing, 2018.
• HRF
  • Images: Budai, Attila, et al. "Robust vessel segmentation in fundus images." International journal of biomedical imaging 2013.1 (2013): 154860.
  • Annotations: Chen, Wenting, et al. "TW-GAN: Topology and width aware GAN for retinal artery/vein classification." Medical Image Analysis 77 (2022): 102340.