ICIP2020 Joint Robust Attention Network

This code is part of the supplementary materials for the IEEE ICIP 2020 for our paper Improving robustness using Joint Attention network for Optical Coherence Tomography Images . The paper has since been accpeted to IEEE ICIP 2020 and will be preseneted in October 2020.

Arxiv Pre-print

https://arxiv.org/abs/2005.08094

IEEE Xplore Digital Library

https://ieeexplore.ieee.org/document/9190742

Citation

@INPROCEEDINGS{9190742,

  author={S. A. {Kamran} and A. {Tavakkoli} and S. L. {Zuckerbrod}},

  booktitle={2020 IEEE International Conference on Image Processing (ICIP)}, 

  title={Improving Robustness Using Joint Attention Network for Detecting Retinal Degeneration From Optical Coherence Tomography Images}, 

  year={2020},

  volume={},

  number={},

  pages={2476-2480},}

Virtual Oral Presentation (Youtube)

Pre-requisite

• Ubuntu 18.04 / Windows 7 or later
• NVIDIA Graphics card

Installation Instruction for Ubuntu

• Download and Install Nvidia Drivers
• Download and Install via Runfile Nvidia Cuda Toolkit 10.0
• Download and Install Nvidia CuDNN 7.6.5 or later
• Install Pip3 and Python3 enviornment

sudo apt-get install pip3 python3-dev

• Install Tensorflow-Gpu version-2.0.0 and Keras version-2.3.1

sudo pip3 install tensorflow-gpu==2.0.0
sudo pip3 install keras==2.3.1

• Install packages from requirements.txt

sudo pip3 -r requirements.txt

Training on Srinivasan2014 Dataset

• Please cite the paper if you use their data

@article{sri2014,
  title={Fully automated detection of diabetic macular edema and dry age-related macular degeneration from optical coherence tomography images},
  author={Srinivasan, Pratul P and Kim, Leo A and Mettu, Priyatham S and Cousins, Scott W and Comer, Grant M and Izatt, Joseph A and Farsiu, Sina},
  journal={Biomedical optics express},
  volume={5},
  number={10},
  pages={3568--3577},
  year={2014},
  publisher={Optical Society of America}
}

Dataset download link for Srnivinasan2014

http://people.duke.edu/~sf59/Srinivasan_BOE_2014_dataset.htm

Folder Structure Rearrangement

• The folder stucture for this dataset needs to be rearranged. We have provided a .py file to rearrange the structure and also, create the necessary directories and sub-directories. Please run the python file data_preprocess_sri2014.py. In cmd/terminal type the below command. Run it as it is or tinker with it, if you wanna change the directory.

python3 data_preprocess_sri2014.py

• The structure will look like this before the rearrangement.

|──2014_BOE_Srinivasan
|  ├──Publication_Dataset
|     ├──AMD1
|     ├──AMD2
|     ├──DME1
|     ├──DME2
|     .....
|     .....
|     └──NORMAL15
|         └──TIFFs
|             └──8bitTIFFs
├── src
├── LICENSE
├── README.md
├── data_preprocess_sri2014.py
├── inference.py
├── requirements.txt
├── test.py
└── train.py

• And it will look like this afterwards.

├── data
|   ├──Srinivasan2014
|       ├──Train
|           ├──AMD
|           ├──DME
|           └──NORMAL
|       ├──Test
|           ├──AMD
|           ├──DME
|           └──NORMAL
├── src
├── LICENSE
├── README.md
├── data_preprocess_sri2014.py
├── inference.py
├── requirements.txt
├── test.py
└── train.py

• Type this in terminal to run the train.py file

python3 train.py --dataset=Srinivasan2014 --datadir=data/Srinivasan2014 --batch=4 --epoch=30 --logdir=optic-net-sri2014-log --snapshot_name=optic-net-sri2014

• There are different flags to choose from. Not all of them are mandatory

   '--batch', type=int, default=8
   '--input_dim', type=int, default=224
   '--datadir', type=str, required=True, help='path/to/data_directory'
   '--epoch', type=int, default=30
   '--logdir', type=str
   '--weights', type=str,default=None, help='Resuming training from previous weights'
   '--model',type=str, default='OpticNet71',help='Pretrained weights for transfer learning',choices=['OpticNet71','ResNet50', 'MobileNetV2']
   '--snapshot_name',type=str, default=None, help='Name the saved snapshot'

License

The code is released under the MIT License, you can read the license file included in the repository for details.