In this paper, we investigate several deep learning models for use in diagnosing diseases from chest X-ray images. We train a simple convolutional neural network (CNN), a ResNet model, a DenseNet model, and an Extended CNN model on a subset of images from the CheXpert dataset of chest X-ray images, both for binary classification of pneumonia and multi-label classification of multiple pathologies. Our Extended CNN model is based on our Simple CNN model but includes features that are not commonly used in X-ray classification, namely multiple image views and electronic health record (EHR) data. Our goal is to see whether these additional features can give researchers another option in which to improve their models. In this study we evaluate the results of our Extended CNN model using accuracy, AUC, and F1 score and compare to our Simple CNN model as the baseline and also to the state-of-the-art models developed by other researchers. Both the ResNet and DenseNet were modified for our needs and did not use pre-trained weights. We find that we are able to achieve and replicate similar performance using these state-of-the-art models with our dataset. The results from our Extended CNN model fell short of those of the state of the art models, however with some modifications we believe there could be some merit with the framework

The full paper in pdf format can be found here
A summary in pptx format can be found here

Example data and structure is found in the /images/ file - this is a drastically reduced dataset from the original CheXpert dataset, to be used to validate only that the provided code runs (specifically the below csvs will work for the multi-label simple CNN/ResNet/DenseNet models).

• /images/restricted_example_train.csv - csv limited to 64 patients with corresponding images available in the /images/CheXpert-v.10-small/train folder
• /images/restricted_example_valid.csv - csv limited to 50 patients with corresponding images available in the /images/CheXpert-v1.0-small/valid folder

Data was obtained from the CheXpert Dataset based on the paper CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison. The code uses the test.csv and valid.csv files to pull the images, so these files can be modified and restricted to the available images for training/testing.

The model building code is found in the /Models/ file.

• Binary Simple CNN: Pneumonia_SimpleCNN.ipynb
• Binary Extended CNN: Pneumonia_ExtendedCNN.ipynb
• Binary ResNet18: Pneumonia_ResNet.ipynb
• Binary DenseNet121: Pneumonia_DenseNet.ipynb
• Multi-label Simple CNN: Multilabel_SimpleCNN.ipynb
• Multi-label Extended CNN: Multilabel_ExtendedCNN.ipynb
• Multi-label ResNet18: Multilabel_ResNet.ipynb
• Multi-label DenseNet121: Multilabel_DenseNet.ipynb

The models were run on Google Colab attached to a Google Drive folder containing the CheXpert dataset. To reference the appropriate data, set the values of the following variables:

IMG_PATH =  #input your image path here
TRAIN_CSV = #input your train.csv file path here
VALID_CSV = #input your valid.csv file path here