stanfordmlgroup / chexaid

CheXaid

This repo contains code accompaning the paper CheXaid: Deep Learning Assistance for Physician Diagnosis of Tuberculosis using Chest X-Rays in Patients with HIV.

This is a deep learning model developed by Stanford Machine Learning Group to help physicians detect Tuberculosis (TB) among patients with HIV infection using their Chest X-rays.

Table of Contents

• Prerequisites
• Data
• Training
• Model Evaluation
• License
• Citing

---

Prerequisites

This code uses python and pytorch. Follow these steps to get started:

1. Create a conda virtual environment:

conda env create -f environment.yml

2. Activate the virtual environment:

conda activate chexaid

3. If using dummy data: create dummy data dataframe:

python create_dummy_df.py

---

Data

By default, the training data directory is set to be dummy/, which has 128 randomly generated images. You can replace it with your own data.

---

Training

To train a model with the default parameters run:

python train.py

Changing trainging parameters

By default, hyperparameters for training are set in util/hypersearch.py. This script includes information about the pretrained model to use for transfer learning (checkpoint paths, number of original classes), data args (task sequence, number of the fold to train and validate on for k-fold cross validation, image transformations), model args (model to be used, list of covariates, loss function), optimizer args and logger args. More details about these can be found in args/.

Alternatively, hyperparameters can be set by adding arguments to the training command:

python train.py \
       --batch_size=64 \
       --eval_pulm=True \
       --num_epochs=10

---

Model Evaluation

In order to test a model that has been trained as above, the test script can be run as:

python test.py \
       --eval_pulm=True \
       --transform_classifier=False \
       --save_cams=True \
       --task_sequence=pulm \
       --ckpt_path=ckpts/new_gt_1549918027265_4CF892/iter_6656_pulm-valid_pulm_tbAUROC_0.79.pth.tar \
       --split=test

Note that the ckpt_path argument should point to the checkpoints of the evaluated model. By default, model checkpoints will be saved following the format: ckpts/new_gt_<timestamp>/<checkpoint name>.pth.tar

---

License

This repository is made publicly available under the MIT License.

---

Citing

If you are using this code, please cite our paper:

@article{rajpurkar2020chexaid,
  title = {CheXaid: deep learning assistance for physician diagnosis of tuberculosis using chest x-rays in patients with HIV},
  author = {Pranav Rajpurkar and Chloe O'Connell and Amit Schechter and Nishit Asnani and Jason Li and Amirhossein Kiani and Robyn L. Ball and Marc Mendelson and Gary Maartens and Daniël J. van Hoving and Rulan Griesel and Andrew Y. Ng and Tom H. Boyles and Matthew P. Lungren},
  journal = {npj Digital Medicine},
  volume = {3},
  number = {1},
  doi = {10.1038/s41746-020-00322-2},
  year = {2020},
  publisher = {Nature Publishing Group}
}

chexaid's Issues

Model weights

Do you have the model weights available?

Could you share your dataset?

Dear authors,
We are building an AI tool for diagnosis of Covid-19, TB, other pneumonia and Normal from both CXR and CT. Our application is free. Therefore, we are collecting as many medical images as we could do.
Can you share your dataset for us?
Best regards.
Linh