Awesome

<center>COVID-DA: Deep Domain Adaptation from <br> Typical Pneumonia to COVID-19</center>

We provide the COVID-DA dataset for domain adaptation from typical pneumonia to COVID-19. The paper is available here.

Dataset

The descriptions for the COVID-DA dataset are presented below. We first provide a link to download the dataset. Next, the data statistics and usage of the dataset will be introduced.

Download

• The dataset in this paper is available here.

Data Composition

To make up this dataset, we collected and integrated the following open-source datasets: https://github.com/ieee8023/covid-chestxray-dataset
https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/data
https://www.kaggle.com/darshan1504/covid19-detection-xray-dataset
https://www.kaggle.com/andrewmvd/convid19-X-rays
https://www.kaggle.com/nabeelsajid917/covid-19-x-ray-10000-images
https://www.kaggle.com/usmantahirkiani/covid19-vs-healthy-xray
https://www.kaggle.com/tarandeep97/covid19-normal-posteroanteriorpaxrays

Data Structure and Statistics

• The data structure:

all_data
└── all_data_pneumonia
|   |
|   ├── train
|   └── val 
|
└── all_data_covid
    |
    |── train
    |── val
    └── test

• Statistics of the dataset are shown as follow:
  
  Pneumonia ("all_data_pneumonia" sub-directory) serves as the source domain and COVID-19 ("all_data_covid" sub-directory) serves as the target domain.

• You can refer to the paper for more details about the dataset.

Usage

• In the directory ./data, there are two .pkl files which record the image lists and its corresponding labels. Specifically, an image and its label is stored in a tuple (image_name, label). "1" denotes class "pneumonia" and class "COVID-19" in source and target domain, respectively, while "0" denotes class "normal". You can read the data list following the below manner:

  • for the source domain (Pneumonia):

        with open('./data/pneumonia_task.pkl', 'rb') as f:
            train_dict = pickle.load(f)
        train_list = train_dict['train_list'] # train sub-directory
        val_list = train_dict['val_list'] # test sub-directory
  

  • for the target domain (COVID-19):

        with open('./data/COVID-19_task.pkl', 'rb') as f:
            train_dict = pickle.load(f)
        train_list_labeled = train_dict['train_list_labeled'] # labeled data (train sub-directory)
        train_list_unlabeled = train_dict['train_list_unlabeled'] # unlabeled data (train sub-directory)
        val_list = train_dict['val_list'] # val sub-directory
        test_list = train_dict['test_list'] # test sub-directory
  

  For convenience, we provide .pkl files for both python 2 and 3, respectively.

• According to the image lists, you can load images using Pillow:

        # e.g., for the source domain (Pneumonia)
        for img_tup in train_list:
            img = PIL.Image.open(os.path.join('all_data/all_data_pneumonia', 'train', img_tup[0])
            label = img_tup[1]
  

Citation

If you find the COVID-DA dataset useful, please cite the following paper:

@article{zhang2020covidda,
    title={COVID-DA: Deep Domain Adaptation from Typical Pneumonia to COVID-19},
    author={Yifan Zhang and Shuaicheng Niu and Zhen Qiu and Ying Wei and Peilin Zhao and Jianhua Yao and Junzhou Huang and Qingyao Wu and Mingkui Tan},
    journal={arXiv},
    year={2020},
}