Right for the Wrong Reason

Official implementation for MICCAI 2023 paper: Right for the Wrong Reason: Can Interpretable ML Techniques Detect Spurious Correlations? Arxiv Paper

Note: For any questions regarding the paper and code, contact Susu Sun with susu.sun@uni-tuebingen.de.

Overview of Experiment settings

We performed a rigorous evaluation of post-hoc explanations and inherently interpretable techniques for the detection of spurious correlations in a medical imaging task.

• We designed three kinds of spurious signals named "Tag", "Hyperintensities" and "Obstruction" and contaminated the positive samples in the train set with ratios of 0%, 20%, 50%, 80%, and 100%.
• We designed two novel metrics named "Confounder Sensitivity" and "Explanation NCC" to evaluate five post-hoc explanation methods and one inherently interpretable model in their ability to detect the spurious signal.

Installation

conda env create -f right_for_wrong.yml
conda activate right_for_wrong

Datasets

We perform experiments on the CheXpert-v1.0-small dataset. It is a small version of the original CheXpert dataset with resized chest x-ray images.

This dataset can be downloaded with Kaggle (https://www.kaggle.com/datasets/ashery/chexpert)

The official CheXpert can be found with (https://stanfordmlgroup.github.io/competitions/chexpert/)

It is also possible to perform on dataset VinDr-CXR (https://vindr.ai/datasets/cxr). You need to download and preprocess the dataset using scripts (https://github.com/ss-sun/Attri-Net/blob/master/data/preprocess_nih_vindr_datasets.py).

How to use the code

Step 1: Generate the contaminated datasets with ratios of 0%, 20%, 50%, 80% and 100% and spurious signals "Tag", "Hyperintensities" and "Obstruction".

• Set the SRC_DATA_ROOT and TGT_DATA_ROOT in data/contaminate_data_settings.py
• Run data/contaminate_data.py by setting the contaminated_dataset, contaminated_class, contamination_type and contamination_scale. E.g python3 ./data/contaminate_data.py --contaminated_dataset chexpert --contaminated_class "Cardiomegaly" --contamination_type tag --contamination_scale 1

Step 2: Train ResNet50 and Attri-Net on the contaminated datasets.

• Set experiment relevant parameters in parser.py.
• Train ResNet50 or Attri-Net by setting the parameter which_model to resnet or attrinet in main.py. E.g. python3 main.py --dataset chexpert --contaminated_class "Cardiomegaly" --contaim_type hyperintensities --contaim_scale 2

Step 3: Generate confounder masks and select the flipped samples to prepare for the evaluation.

• Run create_confounder_mask.py to create ground truth confounder masks for three signals "Tag", "Hyperintensities" and "Obstruction".
• Set the trained models' path in models_dict.py, and run select_flip_prediction_samples.py to select the flipped prediction samples.

Step 4: Perform evaluations of XAI methods.

• Run create_results_dict.py to create a result dictionary for collecting all the evaluation results.
• Evaluate the classification performance with eval_auc.py. Evaluate the explanation NCC and confounder sensitivity with spurious_detection_eval.py.

Results

Qualitative evaluation

Quantitative evaluation

References

If you use any of the code in this repository for your research, please cite as:

 @misc{sun2023right,
      title={Right for the Wrong Reason: Can Interpretable ML Techniques Detect Spurious Correlations?}, 
      author={Susu Sun and Lisa M. Koch and Christian F. Baumgartner},
      year={2023},
      eprint={2307.12344},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}
}