CNN for the Classification of Chest X-ray Scans

---

Disclaimer: This work represents by no means a medical study and is merely an example of what can be done using CNNs in tasks that involve processing images.

In this project, we define and train a CNN to classify chest x-ray scans taken from the dataset by Joseph Paul Cohen and Paul Morrison and Lan Dao.

Two solutions are provided using ResNet34. The first one, implementing the model directly from torchvision.models and the second one, implementing each layer of the model using torch API. We use Cross Entropy Loss and Stochastic Gradient Descent to train the models.

Data is separated between a validation and a training set. So that we can measure the model's performance on data it hasn't seen during training and avoid overfitting.

At the end, this allows us to get an accuracy close to 98% on the validation set.

A Dockerfile is provided. This allows to have a unified runtime environment. After instantiating the docker container, you will have a virtual environment with all the necessary dependencies. Since pytorch:1.5-cuda10.1-cudnn7 container is used, nvidia drivers >= 418 are required.

Run on Linux (Not tested in other OS)

1. Install Docker following the instructions on the link.

2. Install nvidia docker for gpu support from this link

3. Open a terminal

4. Clone this repository and go to the repo folder

   cd covid19_xray_pytoch

5. Build the docker container (this may take several minutes)

   docker build . -t xray_classifier

6. Run the docker container. This will enable a terminal inside the container (user: developer - password: developer).

   docker run -p 8888:8888 -p 6006:6006 -v `pwd`/dev_ws:/usr/src/app/shared -it --rm --gpus all xray_classifier

7. Inside the container, run jupyter lab

   jupyter lab --ip=0.0.0.0 --port=8888 --allow-root

8. Go to the train.ipynb to go through the training process.

9. During training, open a terminal inside jupyter lab by clicking on the '+' button on the top left corner, and type the following to watch tensorboard's dashboard:

   tensorboard --logdir=./runs --bind_all

   This will open tensorboard server, after a couple minutes, go to the following link: localhost:6006 You will be able to see the training process as shown in the figure.