This repository provides proof of concept code for detecting pathology related to COVID-19 from a Sensyne hackathon. The figure below shows the original CT axial slice, the manually annotated ground truth and the automatic delineation of consolidation and ground glass opacity.

Please see the following blog post for more details: https://medium.com/@heaven_spring_cheetah_212/a-weekend-a-virtual-hackathon-and-ml-approaches-to-automate-the-analysis-of-covid-19-lung-ct-5ad4c89fe03c

This code is based on the U-Net implementation from: https://github.com/zhixuhao/unet

1) Dataset

The original dataset is provided at:

https://www.sirm.org/2020/03/28/covid-19-caso-65/

The annotated dataset with annotations are available at from MedSeg

To run the code, download the training images, training masks and test images from the site (tr_im.nii.gz, tr_mask.nii.gz and val_im.nii.gz)

and copy these files into data/covid/ in the code directory.

2) Trained model

The trained model can be downloaded from the the Github release at:

https://github.com/sensynehealth/lung-ct-covid-pathology-detection/releases/download/v0.1/unet_covid_crossval_200.hdf5

This needs to be copied into the main folder.

3) Building the environment and running the code

Running the code in a virtual environment

Create the virtual environment

python3 -m venv venv

pip install -r requirements.txt

To run the code

python3 main.py

(optional) Running the code using a docker container

docker build --tag covidruntime:1.0 .

To run the code

docker run --rm -it -v $(pwd):/code covidruntime:1.0 python3 main.py

Output

The processed images from the test and validation set will appear in demo_figs

4) Usage

The following simple options are available in the main script:

# Options 
# --------------------------------------

train_model = False
validation_size = 0.1
steps_per_epoch = 100
epochs = 200

# --------------------------------------

Selecting train_model = True wil train the model from scratch.