Skip to content

Latest commit

 

History

History
73 lines (48 loc) · 3.41 KB

README.md

File metadata and controls

73 lines (48 loc) · 3.41 KB

Assessing the inter-relationship of prediction tasks

Components of a deep neural networks

This repository will contain the code for the paper

B. Glocker, S. Winzeck. Assessing the inter-relationship of prediction tasks: Implications for algorithmic encoding of protected characteristics and its effect on AI performance. 2021.

Dataset

The CheXpert imaging dataset together with the patient demographic information used in this work can be downloaded from https://stanfordmlgroup.github.io/competitions/chexpert/.

Code

For running the code, we recommend setting up a dedicated Python environment.

Setup Python environment using conda

Create and activate a Python 3 conda environment:

conda create -n pymira python=3
conda activate chexploration

Install PyTorch using conda:

conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

Setup Python environment using virtualenv

Create and activate a Python 3 virtual environment:

virtualenv -p python3 <path_to_envs>/chexploration
source <path_to_envs>/chexploration/bin/activate

Install PyTorch using pip:

pip install torch torchvision

Install additional Python packages:

pip install matplotlib jupyter pandas seaborn pytorch-lightning scikit-learn scikit-image tensorboard tqdm openpyxl

How to use

In order to replicate the results presented in the paper, please follow these steps:

  1. Download the CheXpert dataset, copy the file train.csv to the datafiles folder
  2. Download the CheXpert demographics data, copy the file CHEXPERT DEMO.xlsx to the datafiles folder
  3. Run the notebook chexpert.sample.ipynb to generate the study data
  4. Run the script chexpert.disease.py to train a disease detection model
  5. Run the script chexpert.sex.py to train a sex classification model
  6. Run the script chexpert.race.py to train a race classification model
  7. Run the notebook chexpert.predictions.ipynb to evaluate all three prediction models
  8. Run the notebook chexpert.explorer.ipynb for the unsupervised exploration of feature representations

Additionally, there are scripts chexpert.sex.split.py and chexpert.race.split.py to run SPLIT on the disease detection model. The default setting in all scripts is to train a DenseNet-121 using the training data from all patients. The results for models trained on subgroups only can be produced by changing the path to the datafiles (e.g., using full_sample_train_white.csv and full_sample_val_white.csv instead of full_sample_train.csv and full_sample_val.csv).

Note, the Python scripts also contain code for running the experiments using a ResNet-34 backbone which requires less GPU memory.

Funding sources

This work is supported through funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (Grant Agreement No. 757173, Project MIRA, ERC-2017-STG) and by the UKRI London Medical Imaging & Artificial Intelligence Centre for Value Based Healthcare.

License

This project is licensed under the Apache License 2.0.