ChestXRay-Classification

ChestXRay Disease Diagnosis

About

A Deep Neural Network model for multi-label thorax disease classification on chest X-ray images. The dataset,CheXpert, provides High-Resolution and Low-Resolution images labeled with 14 classes.Our classfication problem consists of only 5 classes as mentioned in the CheXpert Competition.
We experimented with architectures like DenseNet-121, NASNet4, Resnet-152 with similar parameters, activation and loss function.
AUROC and Precision Recall are used as metric for the evaluation of the models.

Architecture/Models

Denset-121 -> classifier (sigmoid activation function per class)

Resnet-151 -> classifier (sigmoid activation function per class)

NAsNetLarge -> classifier (sigmoid activation function per class)

Dataset

CheXpert: https://arxiv.org/abs/1901.07031

We have included a subset of Dataset in this repo for CSE 6250 Evaluation Purpose.

Dataset can be downloaded from here: http://download.cs.stanford.edu/deep/CheXpert-v1.0-small.zip

Note: Please note that you need to subscribe to CheXpert website to download the dataset: https://stanfordmlgroup.github.io/competitions/chexpert/

Models

Model which is submitted to CheXpert competition is available at dir: codalab/src/best_weights_1555982768.7076797.h5

Currently, We are placed 4th (fourth) in the competition.

NASNetLarge model can be downloaded from this link

For quick evaluation of the existing models, please follow below steps:

We have used Anaconda for the project development and testing. requirements_eval.txt lists the required packages if you may choose not to use Anaconda.

Set-up:

conda create -n chexpert python=3.6.5
conda activate chexpert
cd <root directory>
pip install -r requirements_eval.txt

Run the model:

export PYTHONPATH=<Path-To-ChestXRay-Classification>:$PYTHONPATH
python src/test.py

Set Up

pip install -r requirements.txt

Training

export PYTHONPATH=<Path-To-ChestXRay-Classification>:$PYTHONPATH
mkdir <Path-To-ChestXRay-Classification>/out
python src/train.py

Testing

export PYTHONPATH=<Path-To-ChestXRay-Classification>:$PYTHONPATH
python src/test.py --data-dir <Dir Containing CheXpert Data> --model-file-path <Model Weights File Path> --model-type 'DenseNet121'

Results

Below is the model comparison for the 5 classes ('Atelectasis', 'Cardiomegaly', 'Consolidation', 'Edema', 'Pleural Effusion')

Class\Models	DenseNet121	NASNet-Large
Atelectasis	0.808819	0.805774
Cardiomegaly	0.823752	0.800245
Consolidation	0.923713	0.853493
Edema	0.918155	0.925000
Pleural Effusion	0.910326	0.932858
Mean AUROC	0.876953	0.863474

Code Structure

├── README.md
├── codalab
│   ├── CheXpert-v1.0
│   │   └── valid
│   │       └── patient00000
│   │           ├── study1
│   │           │   ├── view1_frontal.jpg
│   │           │   └── view2_lateral.jpg
│   │           └── study2
│   │               └── view1_frontal.jpg
│   ├── src
│   │   ├── best_weights_1555982768.7076797.h5
│   │   ├── codalab_submit.py
│   │   └── models.py
│   ├── src.zip
│   └── valid_image_paths.csv
├── config.ini
├── notebooks
│   ├── dataPrep.ipynb
│   ├── evaluate.ipynb
│   └── inspect_model.ipynb
├── out
├── requirements.txt
├── src
│   ├── augmentations.py
│   ├── callbacks.py
│   ├── generator.py
│   ├── metrics.py
│   ├── models.py
│   ├── test.py
│   ├── train.py
│   └── utils.py
├── test_imgs
│   ├── view1_frontal.jpg
│   ├── view1_frontal2.jpg
│   └── view1_frontal3.jpg
├── train4.out
├── train5.out
└── weights
    ├── best_weights_1555865398.1238055_Apr22_5cls.h5
    └── best_weights_1555982768.7076797.h5

rishabhbhardwaj/ChestXRay-Classification