Y-Net: A Spatiospectral Network for Retinal OCT Segmentation

This is the official source code for the paper "$\Upsilon$-Net: A Spatiospectral Network for Retinal OCT Segmentation" accepted in MICCAI 2022.

Link to arXiv

Authors: A. Farshad*, Y. Yeganeh*, P. Gehlbach, N. Navab

Citation

If you find this code useful in your research then please cite:

@inproceedings{farshad2022_ynet,
	    title={Υ-Net: A Spatiospectral Network for Retinal OCT Segmentation},
	    author={Farshad, Azade and Yeganeh, Yousef and Gehlbach, Peter and Navab, Nassir},
	    booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
	    year={2022},
	    organization={Springer}
	  }

Environment

The requirements.txt file includes the required libraries for this project.

python -m venv ynet
source ./ynet/bin/activate
pip install -r requirements.txt

Datasets Downloading and Preproccesing

Downloads the dataset, creates the required data directories and preprocesses the data:

sh data_download_and_preprocess.sh

Model Evaluation

Evaluate the pre-trained models:

python eval.py

This will report the quantitative comparison between UNet and Y-Net + FFC and save the qualitative comparison to "./figs". To compare the model parameters, set --print_params to True.

Model Training

Train the Y-net + FFC model:

python train.py --dataset [Duke | UMN]

Train the plain Y-net model:

python train.py --model_name y_net_gen --dataset [Duke | UMN]

For UMN, set the number of classes to 2, and the correct data path:

python train.py --dataset UMN --n_classes 2 --image_dir path_to_UMN

Train plain U-Net model:

python train.py --model_name unet --dataset [Duke | UMN]