Covid-19 x-ray image classification

Output:

Builds an image classifier using transfer learning that detects xray images of 'covid-19' vs. 'other' patient's lungs.

Overview:

Model is extremely basic, doesn't use localisation and validation accuracy can vary between 75% to 87% based on several runs.

• Dataset: https://github.com/ieee8023/covid-chestxray-dataset
• Model: resnet18 pre-trained on ImageNet.
• Project modularised and set up to be data agnostic so they be adapted to build other classifiers.
• src/xrayclassifier/trainer/training.py contains main run process

Local model training:

1. Download full dataset to artifacts/dataset in this project repo, with the following paths populated:

   • artifacts/dataset/annotations
   • artifacts/dataset/images
   • artifacts/dataset/metadata.csv

2. Make sure you have docker installed & set current working directory to covid-xray-classifier project.

3. Build docker image in terminal: make Build

4. Run training: make Train

Possible to dos:

• Prediction module.
• Add Jenkinsfile and marathon.json for pipeline and app definition.
• Add data source as container volume to avoid bringing data into project files directly.
• Use bounding box annotations supplied with xray images for more accurate classification.
• Use more advanced models such as image segmentation.

Some useful code borrowed from: https://github.com/bentrevett/pytorch-image-classification/blob/master/5%20-%20ResNet.ipynb