FL-COVID-19

Implementation of our manuscript of Federated deep learning for detecting COVID-19 lung abnormalities in CT: A privacy-preserving multinational validation study

by Qi Dou#, Tiffany Y So, Meirui Jiang, Quande Liu, VARUT Vardhanabhuti, Georgios Kaissis, Zeju Li, Weixin Si, Heather Lee, Kevin Yu, Zuxin Feng, Li Dong, Egon Burian, Friederike Jungmann, Rickmer Braren, Marcus Makowski, Bernhard Kainz, Daniel Rueckert, Ben Glocker#, Simon CH Yu#, Pheng Ann Heng#

For any inquiry, please contact Dr. Qi Dou ([email protected]).

Source code

The source code enables both training and testing of our federated deep learning method for abnormalities detection in COVID-19 lung CT. We use improving RetinaNet [1] as our backbone.

Installation

1. We recommend using conda to setup the environment, See the requirements.yaml for environment configuration

   If there is no conda installed on your PC, please find the installers from https://www.anaconda.com/products/individual

   If you have already installed conda, please change the directory to the same level as environment.yaml and use the following command to setup environment quickly.

   conda env create -f environment.yaml

   After environment creation, please use the following command to activate your environment

   conda activate flcovid

2. Run the script setup-env.sh to compile Cython code first.

   bash setup-env.sh

3. Put the download demo data and model at the same level of folder 'fl_covid'. Make sure put the data folder and model folder in right place, please refer to the File structure at the bottom.

Usage with examples

Train

To start traning, please run the train.sh.

cd FL-COVID
bash train.sh

You can modify and run your own command according to the template in train.sh

python3 ./fl_covid/bin/train_fed.py --batch-size=6 --steps=100 --gpu=0 --epochs=30 --snapshot-path=../fl_covid_model_and_log/

• --batch-size

  Size of the training batches

• --steps

  Number of steps per epoch

• --epochs

  Number of epochs to train

• --gpu

  Id of the GPU to use (as reported by nvidia-smi)

• --snapshot-path

  Path to save the snapshots of models and training log during training

Test

To run inference with our model, please run test.sh. Inference will be ran on external site1 by default.

cd FL-COVID
bash test.sh

To choose which site to be validated, please change --site in shell script. Options please refer to the below explanation .

You may specify the --save-path to change the results saving directory , default folder is named 'output' at the same directory level as the code folder.

# Run inference and saving the visual and statistical results
python3 ./fl_covid/bin/evaluate_overall_patient_wise.py --model=./model/model.h5  --gpu=0 --save-path=./output/ --save-result=1  --get_predicted_bbox --site=external1

# Print the statistical analysis
python3 ./fl_covid/bin/evaluate_overall_patient_wise.py --model=./model/model.h5  --gpu=0 --save-path=./output/ --save-result=0 --verbose --site=external1

• --model

  Path to the model for evaluation

• --save-path

  Path to save the visualization results and statistical intermediate results

• --save-result

  ‘1’ for both saving visual and statistical results, '0' for only statistics.

• --get_predicted_bbox

  To be used for first time evaluation, saving intermediate results for following statistics.

• --site

  Determine which site to be evaluated. Options are [ internal | external1 | external2 | external3 ]

Input and Output files

The system take input from csv files that saving the data information, each line represents a scan image with label and unique path.

Outputs are statistical results saved in npy format and visual results saved in nii.gz format under the directory FL-COVID/output

For visualizing the nii.gz results, we recommand using ITK-SNAP (http://www.itksnap.org/pmwiki/pmwiki.php?n=Downloads.SNAP3) to visualize them.

File structure

For code and demo data, structures listed below:

|--FL-COVID  
    |--fl_covid (source code folder)  
    |--data (demo date folder)  
        |--internal  
            |--internal_1  
                |--h5  
                |--h5_normalize  
            |--internal_2  
                ...  
            |--internal_3  
                ...  
            |--internal_test  
                |--h5  
                |--h5_normalize  
                |--lung_seg_h5  
        |--external1  
            |--h5  
            |--h5_normalize  
            |--lung_seg_h5  
        |--external2  
            ...  
        |--external3  
            ...  
    |--model (pre-trained model folder)
        model.h5
    |--LICENSE  
    |--README.md  
    |--setup-env.sh  
    |--train.sh  
    |--test.sh  

References

[1] Zlocha, Martin, Qi Dou, and Ben Glocker. "Improving RetinaNet for CT lesion detection with dense masks from weak RECIST labels." In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 402-410. Springer, Cham, 2019.