Diagnosing Respiratory Disease from Chest X-Rays using Computer Vision

Problem

We seek to classify common respiratory diseases (viral pneumonia, lung opacity, COVID, and normal) from chest x-ray images via a survey of supervised and self-supervised methods.

Approach

Structure

As different individuals worked on the various approaches, each of the notebooks under models is self-sustained. We have also provided a requirements.txt which contains all of the necessary dependencies. Each of the models can be run on Google Colab. While the CNN models (ResNet18 and VGG16) can train on a cpu, EVAE and MAE require a gpu for training.

diagnosing_respiratory_disease_cv
│   README.md
└───requirements.txt
└───data
│   └───data_origination.ipynb
└───models
│   └───baseline_models.ipynb
│   └───resnet_vgg.ipynb
│   └───evae.ipynb
│   └───mae_vit.ipynb
└───ignore

Instructions For Use

1. baseline_models.ipynb

• Set the path to the COVID-19 Radiography Dataset in the first code block under section 1. Read / Format Data.
  • The pathway should include everything except the name of the folder where the data is stored
  • An example of a correctly formatted data path is given in the notebook.

2. resnet_vgg.ipynb

• Select the desired options for classification task, model type, and training task
  • Classification options: ['multiclass', 'binary']
  • Model options: ['resnet', 'vgg']
  • Training options: ['finetune', 'scratch']
• Format data into folders consisting of train/val/test splits
  • In the notebook, this folder is called 'train_val_test', and contains a 60:30:10 split for all raw images for each class

3. evae.ipynb

• Amend pathfiles to data
• Requires pre-trained ResNet18 and VGG16 (loads in saved weights)
  • Revision of filepaths for these models is sufficient for implementation to run
• Training recommended on GPU; depending on performance 20 epochs can take 3+ hours

4. mae_vit.ipynb

• Amend pathfiles to data
• Training recommended on GPU; depending on performance 20 epochs can take 3+ hours