Skip to content

sumedhasingla/ExplainingBBSmoothly

Repository files navigation

Official Tensorflow implementation of papers:

Explaining the Black-box Smoothly- A Counterfactual Approach Paper

Using Causal Analysis for Conceptual Deep Learning Explanation Paper

Usage

Train baseline black-box classification model

  1. Download StanfordCheXpert and MIMIC-CXR dataset

  2. Train a classifier. Skip this step if you have a pretrained classifier.

python Step_1_train_classifier.py --config 'configs/Step_1_StanfordCheXpert_Classifier_256.yaml'
  1. Save the output of the trained classifier. User can choose to save the predictions on test set of the same dataset or some different dataset. Set the config appropriately. We used this file to save the prediction on MIMIC-CXR dataset.
python Step_1_test_classifier.py --config 'configs/Step_1_StanfordCheXpert_Classifier_256.yaml'

Generate counterfactual visual explanation

  1. Process the output of the classifier and create input for Explanation model by discretizing the posterior probability. We used this file to create training dataset for Explainer for Pleural Effusion, Cardiomegaly and Edema.
./notebooks/Step_1_Test_Classifier_&_Create_Input_For_Explainer.ipynb
  1. Train a Segmentation Network. We train a lung segmentation network on JSRT dataset
python Step_2_train_segmentation.py --config 'configs/Step_2_JSRT_Segmentation_256.yaml'
  1. Train a Object Detector for pacemaker and hardware. The code is borrowed from: Faster_RCNN_TensorFlow
python Step_3_train_fast_RCNN.py --config 'configs/Step_3_MIMIC_OD_Pacemaker.yaml'
  1. Trainer explainer model.
python Step_4_train_explainer.py --config 'configs/Step_4_MIMIC_Explainer_256_Pleural_Effusion.yaml'
  1. Explore the trained Explanation model and see qualitative results.

Cardiomegaly

python Step_4_test_explainer.py --config 'configs/Step_4_MIMIC_Explainer_256_Pleural_Effusion.yaml'
./notebooks/Step_4_Visualize_and_Quantify_Explainer_Results_Cardiomegaly.ipynb

Includes

  • Pretrained segmentation network for lung and heart. The segmentation network is trained on JSTR dataset and hence is not perfect on MIMIC dataset. Improving the segmentation network is work in progress.
  • Code to compute Cardio Thoracic Ratio (CTR) using lung + heart segmentation.
  • Code to create an animated Gif to better visualize the results, on a given subject.

Pleural Effusion

python Step_4_test_explainer.py --config 'configs/Step_4_MIMIC_Explainer_256_Pleural_Effusion.yaml'
./notebooks/Step_4_Visualize_and_Quantify_Explainer_Results_Pleural_Effusion.ipynb

Includes:

  • Pre-trained object detector for healthy/blunt costophrenic recess.

Generate conceptual explanation

  1. Dissect the trained classification model to identify hidden units which are relevant for a given concept
./notebooks/Step_5_Find_Concept_Units_From_MAX_Activations.ipynb
python Step_1_test_classifier.py --config 'configs/Step_1_StanfordCheXpert_Classifier_256_temp.yaml'
Rscript --vanilla Step_5_Lasso_Regression.r

  1. Use this notebook to visualize the activation regions for a given hidden unit
./notebooks/Step_6_Visualize_Hidden_Units.ipynb

  1. Use the counterfactual explanation to compute the causal in-direct effect associated with a given set of concept-units
./notebooks/Step_7_Causal_Indirect_Effect.ipynb
python Step_4_test_explainer.py --config 'configs/Step_4_MIMIC_Explainer_256_Pleural_Effusion.yaml'

Cite

@article{
ExplainingBBSmoothly,
  author    = {Sumedha Singla and
               Brian Pollack and
               Stephen Wallace and
               Kayhan Batmanghelich},
  title     = {Explaining the Black-box Smoothly- {A} Counterfactual Approach},
  volume    = {abs/2101.04230},
  year      = {2021},
  url       = {https://arxiv.org/abs/2101.04230}
}
@article {
Singla2021Explanation,
	Title = {Using Causal Analysis for Conceptual Deep Learning Explanation},
	Author = {Singla, Sumedha and Wallace, Stephen and Triantafillou, Sofia and Batmanghelich, Kayhan},
	DOI = {10.1007/978-3-030-87199-4_49},
	Volume = {12903},
	Year = {2021},
	Journal = {Medical image computing and computer-assisted intervention : MICCAI ... International Conference on Medical Image Computing and Computer-Assisted Intervention},
	Pages = {519—528}
}

About

Explaining the Black-box Smoothly - A counterfactual Approac; Using Causal Analysis for Conceptual Deep Learning Explanation

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages