CheXNet

A Python3 (TensorFlow) reimplementation of CheXNet paper - Classification and Localization of Thoracic Diseases

Prerequisites

• Python 3.8+

• TensorFlow 2.5+

• All the specified requirements in the text file

Deployment

This model is deployed with a REST API at http://chexnet-1.public.dagshubusercontent.com:8080/invocations. A streamlit client for the deployed model is provided in deployment/src/streamlit_client.py

To deploy this model yourself, follow the instructions in deployment/README.md.

Training

1. Clone this repository.

2. Install requirements.txt using pip install -r requirements.txt.

3. Use DVC to pull the files that are stored on the DagsHub remote storage by running dvc pull

4. Modify the code as you wish.

5. Run dvc repro to run the pipeline and train the model.

The resultant heatmaps from the evaluation stages should look something like 👇

The ChestX-ray14 Dataset

[comment]: The 45GB dataset comprises 112,120 frontal-view chest X-ray images organized 14 different folders, based on individual tarballs. There is a master list of images with target labels, image IDs, patient IDs alongside additional patient metadata. The dataset is available on the Kaggle platform.

1GB-Dataset

To use a smaller, subsampled dataset, run the following commands from the repository's root directory:

git update-index --assume-unchanged data_labeling/data.dvc
curl https://dagshub.com/nirbarazida/CheXNet/raw/d74cccbd0957c41dac7560565d2b7b4df3c0d195/data_labeling/data.dvc -o data_labeling/data.dvc

Acknowledgements

• Data

• License: CC BY 4.0

• Citation

Note: *If you are adding/removing/moving files to different directories, it can affect the DVC pipeline, and therefore the *dvc repro command might not run properly.