Detecting the Causes of Gender Bias in Chest X-ray Diagnosis

Code repository for 2023 MICCAI workshop FAIMI paper: Are Sex-based Physiological Differences the Cause of Gender Bias for Chest X-ray Diagnosis?

Enviroment Set-up

The code repo should be functioning well under python==3.9.11.

Create a new conda enviroment:
conda create -n detecting_causes python=3.9.11

Install the packages in requirements.txt:
conda install -f requirements.txt

Or: (prob work better)
conda env create -f env.yml

Datasets

Both datasets are available online.
For a easy way to download it:

• NIH: link. Download it and unzip it based on its instructions.
• CheXpert: link.
  • A smaller version on Kaggle: link: resized image. Which does not matter that much for this work, as we will resize it to 224 x 224 when preprocessing. Need to mention here that resizing the image will definitely influence the prediction ability; however, in this work, we care more about the performance gap between groups rather than getting better performance.
  • Notice that we only use the frontal images (include PA and AP view position) in this work.

Preprocessing

1. mata data

Mata data has already been processed, please refer to chexpert.sample.allrace.csv (CheXpert) and Data_Entry_2017_v2020_clean_cplit.csv (NIH).

2. images

run the following command to pre-process the images (resize).
python3 ./preprocess/preprocimgs.py -p {your dataset path}

Training

Command example for training NIH dataset on label Pneumothorax with resampling strategy with random state from 0 to 10, 0, 50 and 100 female percentage in training:

First, go to the predction folder:
cd ./prediction/

python3 ./disease_prediction.py -s NIH -d Pneumothorax -f 0 50 100 -n 1 -r 0-10 -p {your dataset folder path}

Train on CheXpert:
python3 ./disease_prediction.py -s chexpert -d Pneumothorax -f 0 50 100 -n 1 -r 0-10 -p {your dataset folder path}

Train on different disease labels sequently:
python3 ./disease_prediction.py -s NIH -d Pneumothorax Pneumonia Cadiomegaly -f 0 50 100 -n 1 -r 0-10 -p {your dataset folder path}

Details about the other hyper-parameters could be found in the same py file.

Sampling actually takes quite a time: #TODO: prepare the splits for different random states as csv in ./datafiles/

Plotting the results

For plotting all the disease after re-store the results into csv file:
python3 ./analysis/plotting.py -p {your run result dir path}

If you train the model with changed hyper-parameters rather then the required 5 (-s -d -f -n -r), you need to add another hyperparameter mannually as -f. For example:
python3 ./analysis/plotting.py -p {your run result dir path} -f -image_size1024
when set --image_size 1024 during training.

(Well, you need to edit the image a bit to make it like this;))

#TODO When specify dataset as ['NIH','chexpert'], the plotting has index errors. Fix it!!

Detecting the Causes

1. Representation Bias (Imbalance Dataset)

See above Plotting the results.

2. Feature Bias (Breast shadows)

To run the experiments with cropped images:
python3 ./prediction/disease_prediction.py -s NIH -d Pneumothorax -f 0 50 100 -n 1 -r 0-10 --crop 0.6 -p {your dataset folder path}

Notice that this part of experiment does not apply to CheXpert.

3. Label Errors

3.1 change the sampling strategies

• Without prioritizing the diseased one: #TODO
• Without sampling or sample more than one scans per patient: change the -n hyper-parameter:
  python3 ./prediction/disease_prediction.py -s NIH -d Pneumothorax -f 0 50 100 -n None -r 0-10 -p {your dataset folder path}
• Change the prevalence setting: --prevalence_setting total or --prevalence_setting equal

3.2 cross sampling strategies experiments: run which py

1. Save model parameters: python3 ./prediction/disease_prediction.py -s NIH -d Pneumothorax -f 0 50 100 -n None -r 0-10 -p {your dataset folder path} --save_model True
2. Run the cross dataset inference: python3 ./analysis/cross_ds_inference.py -d Penumothorax --run_dir {your run dir} --data_dir {your dataset dir}