DRAGON Preprocessing

This repository contains the preprocessing scripts for the DRAGON challenge.

If you are using this codebase or some part of it, please cite the following article: PENDING

BibTeX:

PENDING

Installation

dragon_prep can be pip-installed:

pip install dragon_prep

Alternatively, it can be installed from source:

git clone https://github.com/DIAGNijmegen/dragon_prep
cd dragon_prep
pip install -e .

The Docker can be built after cloning the repository. The anonymisation code is not included due to privacy concerns, so you have to uncomment copying and installing the diag-radiology-report-anonymizer. The unmodified version is included to reflect the exact code used to prepare the DRAGON challenge resources.

git clone https://github.com/DIAGNijmegen/dragon_prep
cd dragon_prep
nano Dockerfile  # uncomment copying and installing the diag-radiology-report-anonymizer
./build.sh

If ran successfully, this results in the Docker container named dragon_prep:latest.

Resources

The preprocessing scripts for the synthetic datasets can be found in src/dragon_prep and are the script called Task1xx_Example_yy.py. The preprocessing scripts for the datasets used in the test leaderboard for the DRAGON challenge can be found in src/dragon_prep and are the script called Task0xx_yy.py. The datasets for the validation leaderboard are derived from the development data, using the src/dragon_prep/make_debug_splits.py script. For the DRAGON challenge, all datasets were preprocessed using the preprocess.sh script.

Usage

The synthetic datasets can be generated with any number of samples.

After installing the dragon_prep module:

python src/dragon_prep/Task101_Example_sl_bin_clf.py \
    --output_dir=./output \
    --num_examples={set any number you like}

Or, using the Docker container:

docker run --rm -it \
    -v /path/to/store/data:/output \
    dragon_prep:latest python /opt/app/dragon_prep/src/dragon_prep/Task101_Example_sl_bin_clf.py \
        --num_examples={set any number you like}


# ... same for Task102_Example_sl_mc_clf.py, Task104_Example_ml_bin_clf.py, Task105_Example_ml_mc_clf.py, Task106_Example_sl_reg.py, Task107_Example_ml_reg.py, Task108_Example_sl_ner.py, Task109_Example_ml_ner.py
# for Task103_Example_mednli.py, setting the number of examples is not supported

The preprocessing scripts for the tasks in the DRAGON benchmark are included for transparancy and to provide building blocks to process your own data. To run the end-to-end script using your own data, you can turn off the anonymisation functionality:

prepare_for_anon(df=df, output_dir=output_dir, task_name=task_name, tag_phi=False, apply_hips=False)

Managed By

Diagnostic Image Analysis Group, Radboud University Medical Center, Nijmegen, The Netherlands

Contact Information

Joeran Bosma: Joeran.Bosma@radboudumc.nl