• Model files were too large hence they were removed from code submission
• Model Files are committed at :
  • classification/model_originial/densenet.pth -- Original CheXpert dataset
  • classification/model_localized/densenet.pth -- Localized dataset
  • localization/code/trained_model.hdf5 -- Pre-trained model
  • localization/code/model.009.hdf5 -- Self-trained model

Localization contains code for extracting localized lung region images from the input chest X-ray images.

• We have used this [U-Net implementation] (https://github.com/imlab-uiip/lung-segmentation-2d) and modified their code where needed for data load or prediction.
• We have also used their pre-trained model for prediction purposes.

• Setup the conda environment using conda env create <environment yml file>
• To run on CPU, use code/env_cpu/environment.yml as environment yml file
• To run on GPU, use code/env_gpu/environment.yml as environment yml file

• JSRT dataset for 247 chest-Xray images
• Corresponding left and right lung region masks from SCR database
• Chexpert images to predict the lung region masks using the model
• Above links should be used to download the data and provide the path to downloaded data in the code before trigerring a run

• Run code/preprocess.py :
  • to perform histrogram equalization on JSRT chest-Xray images
  • to combine left and right lung masks into single image
  • replace jsrt_path variable with JSRT dataset path
  • replace left_lungs_mask_path with path of left lung mask images from SCR database
  • replace right_lungs_mask_path with path of right lung mask images from SCR database
  • replace preprocess_output_path variable with pre-processing output directory. This directory should be created before running the code.
• Run code/train_model.py to train the model for generating lung masks using U-Net implementation with:
  • preprocessed JSRT chest-Xray images as X (input vector)
  • preprocessed single image for left and right lung masks as Y (output vector)
  • To run the file, replace path variable with JSRT dataset path
• Run code/inference.py to use the model for generating lung masks from Chexpert Images. To run the file:
  • replace path variable with Chexpert dataset path
  • set batch variable value as 'train' or 'valid'
  • best perfoming model is commited as file named model.009.hdf5
  • pre-trained model is commited as file named trained_model.hdf5

Classification contains code for predicting diseases from images with labels.

• Setup the conda environment using conda env create <environment yml file>
• Where environment file is, code/environment.yml

• Chexpert overlay dataset containing 28,929 chest X-ray images
• Chexpert original dataset corresponding to the above 28,929 overlay images
• Above links should be used to download the data and provide the path to downloaded data in the code before trigerring a run

• Run code/python etl_chexpert_data.py -h to get full detailed command usage: etl_chexpert_data.py [-h] -c CSV_PATH -p PREFIX_PATH -d DEST_PATH -o OVERLAY

divide the data in train validate and test

optional arguments: -h, --help show this help message and exit -c CSV_PATH Path to file containing file name and labels -p PREFIX_PATH Path to directory containing image dataset -d DEST_PATH Path to output directory -o OVERLAY Is overlay images Y for yes N for no!

• Run code/python train_densenet.py -h usage: train_densenet.py [-h] -p PREFIX_PATH

train densenet 121 with 9 epocs and batch size of 50

optional arguments: -h, --help show this help message and exit -p PREFIX_PATH Path to directory containing image dataset