A model for classifying different Retinal Diseases using Deep Learning from Optical Coherence Tomography Images. This code is part of the supplementary materials for the IEEE ICMLA 2019 for our paper Optic-net: A Novel Convolutional Neural Network for Diagnosis of Retinal Diseases from Optical Tomography Images. The paper has since been accpeted to IEEE ICMLA 2019 and will be preseneted in December 2019.
- Ubuntu 18.04 / Windows 7 or later
- NVIDIA Graphics card
- Download and Install Nvidia Drivers
- Download and Install via Runfile Nvidia Cuda Toolkit 9.0
- Download and Install Nvidia CuDNN 7.1 or later
- Install Pip3 and Python3 enviornment
sudo apt-get install pip3 python3-dev
- Install Tensorflow-Gpu version-1.12 and Keras version-2.2.4
sudo pip3 install tensorflow-gpu==1.12.1
sudo pip3 install keras==2.2.4
- N.B. If you don't have GPU, you can also build Tensorflow-CPU
sudo pip3 install tensorflow==1.12.1
sudo pip3 install keras==2.2.4
- Install packages from requirements.txt
sudo pip3 -r requirements.txt
- Download and Install Nvidia Drivers
- Download and Install Nvidia Cuda Toolkit 9.0
- Download and Install Nvidia CuDNN 7.1 or later
- Install Pip and Python3 environemnt from Web link
- Install Tensorflow-Gpu version-1.12 and Keras version-2.2.4
sudo pip3 install tensorflow-gpu==1.12.1
sudo pip3 install keras==2.2.4
- N.B. If you don't have GPU, you can also build Tensorflow-CPU
sudo pip3 install tensorflow==1.12.1
sudo pip3 install keras==2.2.4
- Install packages from requirements.txt
sudo pip3 -r requirements.txt
- Download the weight file for model trained on Kermany2018 Link
- Download the weight file for model trained on Srinivasan2014 Link
- Run the inference.py file with the below command
python3 inference.py --imgpath='location of the testing image(single file)' --weights='location to the .h5 file' --dataset='Srinivasan2014 or Kermany2018'
- Please cite the paper if you use their data
@article{kermany2018identifying,
title={Identifying medical diagnoses and treatable diseases by image-based deep learning},
author={Kermany, Daniel S and Goldbaum, Michael and Cai, Wenjia and Valentim, Carolina CS and Liang, Huiying and Baxter, Sally L and McKeown, Alex and Yang, Ge and Wu, Xiaokang and Yan, Fangbing and others},
journal={Cell},
volume={172},
number={5},
pages={1122--1131},
year={2018},
publisher={Elsevier}
}
https://data.mendeley.com/datasets/rscbjbr9sj/3
- Folder structure for training given below. Please make sure it matches with your local repository.
├── data
| ├──OCT2017
| ├──train
| ├──CNV
| ├──DME
| ├──DRUSEN
| └──NORMAL
| ├──test
| ├──CNV
| ├──DME
| ├──DRUSEN
| └──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=Kermany2018 --datadir=data/OCT2017 --batch=4 --epoch=30 --logdir=optic-net-oct2017-log --snapshot_name=optic-net-oct2017
- There are different flags to choose from. Not all of them are mandatory
'--dataset', type=str, required=True, help='Choosing between 2 OCT datasets', choices=['Srinivasan2014','Kermany2018']
'--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'
'--snapshot_name',type=str, default=None, help='Name the saved snapshot'
- 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}
}
http://people.duke.edu/~sf59/Srinivasan_BOE_2014_dataset.htm
- The folder stucture for this dataset needs to be rearranged. We have provided a notebook to rearrange the structure and also, create the necessary directories and sub-directories. Please run the ipython notebook data_preprocess_sri2014.ipynb. In cmd/terminal type the below command and then open the notebook file. Run it as it is or tinker with it, if you wanna change the directory.
jupyter-notebook
- 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
'--dataset', type=str, required=True, help='Choosing between 2 OCT datasets', choices=['Srinivasan2014','Kermany2018']
'--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'
'--snapshot_name',type=str, default=None, help='Name the saved snapshot'
The code is released under the MIT License, you can read the license file included in the repository for details.