----A Pytorch implementation of Noise-Powered Disentangled Representation for Unsupervised Speckle Reduction of Optical Coherence Tomography Images.

data

• train_valid: the folder to store data for training and validation.
• test: the folder to store test data.

dataset

• dataset.py: python script to build pytorch Dataset and DataLoader.

models

• base_model.py: the father class implementing the network building, setup input, forward computation, backpropagation, network saving and loading, learning rate schedulers, and visualization of losses and metrics.
• drgan_model.py: containing the forward and backward process of our proposed model.
• drgan_nets.py: the implementation of our proposed network achitectures, such as Encoder, Decoder etc.

run

• trainer.py: a basic template python file for training from scratch, or resuming training, and validation the Model.
• tester.py: a basic template python file for testing the Model.

lib

• help_functions.py: the python file can be used to store and modify the model initilization strategies and optimizer scheduler settings.
• metrics.py: the python file can be used to store and modify the evaluation metrics.
• visualizer.py: the python file can be used for visualization of the losses and images.

main.py: the script for running the code (train/validation/test).

• if you want to train the model from scratch, run

python main.py --mode train --start_epoch 1

• if you want to resume the training process, run

python main.py --mode train --start_epoch epoch_you_want_to_resume

• if you want to test the model, run

python main.py --mode test --load_epoch parameters_epoch_you_want_to_test

configs.py: the python file can be used to store and modify the hyper-parameters for training, validation and testing process.

• Clone this repo:

git clone https://github.com/tsmotlp/DRGAN-OCT

cd DRGAN

• Install PyTorch 1.0+ and other dependencies (e.g., Pillow, torchvision, visdom)

if you want to train and test DRGAN on your own data, your just need to:

• prepare you own train, validation and test data into folder data.
• modify the hyper-parameters in configs.py to make it suitable for your own data.
• follow the instructions of main.py.
• To view training results and loss plots, run python -m visdom.server and click the URL http://localhost:8097.

If you use this code for your research, please cite our paper.

@ARTICLE{9296372,
  author={Y. {Huang} and W. {Xia} and Z. {Lu} and Y. {Liu} and H. {Chen} and J. {Zhou} and L. {Fang} and Y. {Zhang}},
  journal={IEEE Transactions on Medical Imaging}, 
  title={Noise-Powered Disentangled Representation for Unsupervised Speckle Reduction of Optical Coherence Tomography Images}, 
  year={2020},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TMI.2020.3045207}}

if you have any questions, please email to: [email protected].