Distilling BlackBox to Interpretable models for Efficient Transfer Learning

Project Page | Paper | arXiv | Workshop paper (IMLH) | Video | Poster | Slides

Official PyTorch implementation of the paper
Distilling BlackBox to Interpretable models for Efficient Transfer Learning
Shantanu Ghosh¹, Ke Yu², Kayhan Batmanghelich^1
1BU ECE, ²Pitt ISP
In MICCAI, 2023, Early accept, top ~ 14%

Table of Contents

1. Objective
2. Environment setup
3. Downloading data
   • (b) Downloading MIMIC-CXR
   • (b) Downloading Stanford-CXR
4. Preprocessing MIMIC-CXR
5. Training MoIE-CXR
   • (a) Running MoIE-CXR
   • (b) FOLs for MoIE-CXR
   • (c) Checkpoints MoIE-CXR
6. Finetune for Stanford-CXR
   • (a) Finetuned checkpoints of Stanford-CXR
   • (b) Generating dataset for Semi-supervised learning (SSL)
   • (c) Finetuning MoIE-CXR for transfer learning
   • (d) Generate FOLs for the samples of Stanford-CXR
7. Suggestions
8. How to Cite
9. License and copyright
10. Contact

Objective

In this paper, we aim to apply MoIE (in ICML, 2023) to Chest-X-rays (MIMIC-CXR). Due to class imbalance in large CXR datasets, early experts in MoIE tend to cover all samples with disease present while ignoring disease subgroups and pathological heterogeneity. We address this problem by estimating the class-stratified coverage from the total data coverage. Then we utilize our method for efficient transfer learning to a target domain (Stanford-CXR). Illustration of our method is summarized below:

Environment setup

conda env create --name python_3_7_rtx_6000 -f environment.yml
conda activate python_3_7_rtx_6000

Downloading data

After downloading data from the below links, search for --data-root variable in the codebase and replace the appropriate paths for all the different datasets. Also search for /ocean/projects/asc170022p/shg121/PhD/ICLR-2022 and replace with appropriate paths.

(a) Downloading MIMIC-CXR

• MIMIC-CXR
• RadGraph
• NVIDIA Annotation
• Chest ImaGenome

For more details please follow the AGXNet Repository.

(b) Downloading Stanford-CXR

For downloading, Stanford-CXR, visit the link.

Preprocessing MIMIC-CXR

To preprocess MIMIC-CXR to generate the concepts from the reports, follow the following steps sequentially:

1. To generate itemized RadGraph examples, run:

python ./src/codebase/data_preprocessing/mimic-cxr/miccai-main/preprocessing/radgraph_itemized.py

2. Run ./preprocessing/radgraph_parsed.py To parse RadGraph relations, run:

python ./src/codebase/data_preprocessing/mimic-cxr/miccai-main/preprocessing/radgraph_parsed.py

3. To create adjacency matrix that represents the relations between anatomical landmarks and observations mentioned in radiology reports, run:

python ./src/codebase/data_preprocessing/mimic-cxr/miccai-main/preprocessing/adj_matrix.py

Step 3 will be the concepts for training MoIE-CXR. For more details please follow the AGXNet Repository.

Also, remove the disease label to be classified from the concepts. For example, to classify Pneumonia (disease label), Pneumonia will appear in the concepts produced in Step 3 as we are extracting anatomies and observations using Rad-graph. This is redundant. So in this case, manually remove Pneumonia from the concepts.

If you don't want to run the pre-processing steps for MIMIC-CXR for the radgraph files to get the concepts, please refer to the following paths directly (these will be the outputs of the above 3 steps) and place the files in respective folders as indicated in code for training the blackbox, concept predictor (t) and experts in MoIE-CXR:

Variable	Description	Paths
--radgraph-adj-mtx-pickle-file	radgraph adjacent matrix landmark - observation	landmark_observation_adj_mtx_v2.pickle
--radgraph-sids-npy-file	radgraph study ids	landmark_observation_sids_v2.npy
--radgraph-adj-mtx-npy-file	radgraph adjacent matrix landmark - observation	landmark_observation_adj_mtx_v2.npy
--nvidia-bounding-box-file	bounding boxes annotated for pneumonia and pneumothorax	mimic-cxr-annotation.csv
--imagenome-radgraph-landmark-mapping-file	Landmark mapping between ImaGenome and RadGraph	landmark_mapping.json

Training MoIE-CXR

(a) Running MoIE-CXR

• As a first step find and replace the project path /ocean/projects/asc170022p/shg121/PhD/ICLR-2022 from the whole codebase with appropriate path.
• Go the ./scripts/scripts_mimic_cxr folder and get the training scripts. For all the diseases in for MIMIC-CXR, one script is included in the respective folder with proper instructions to run 1) Blackbox 2) projection (t) 3) interpretable model (g) 4) residual (r). For example, to run cardiomegaly for MIMIC-CXR, refer to the file ./scripts/scripts_mimic_cxr/cardiomegaly/cardiomegaly.sh
• The naming convention and the paths to be replaced is mentioned in the script. Follow them carefully
• Run them sequentially.
• Also, after test_t_mimic_cxr.py, sub-select the concepts whose validation auroc >= 0.7 using the notebooks for each disease in the folder ./src/codebase/jupyter_notebook/MIMIC-CXR.
• Also, after training and testing each expert in MoIE-CXR, FOLs_mimic_cxr_main.py file is responsible for generating instance specific FOL. This file uses ./src/codebase/MIMIC_CXR/paths_mimic_cxr.json file where we keep all the paths and filenames of the checkpoints of Blackbox (bb), concept predictor (t), explainer ( g), and residual (r) for each disease. Replace those paths and filenames with the appropriate ones based on the experiments.

(b) FOLs for MoIE-CXR

We have included the instance-specific explanations per expert for each disease of MIMIC-CXR for 0^th seed in the folder ./explanations. For example, to see the FOLs for cardiomegaly, refer to actual_explanations column of ./explanations.

(c) Checkpoints MoIE-CXR

For the checkpoints of 0^th seed of the pretrained blackboxes and concept banks of different diseases of MoIE-CXR, refer below:

Blackbox (BB)	Concept predictor (t) / Concept banks
Cardiomegaly	Cardiomegaly
Effusion	Effusion
Edema	Edema
Pneumonia	Pneumonia
Pneumothorax	Pneumothorax

Finetune for Stanford-CXR

(a) Finetuned checkpoints of Stanford-CXR

Refer to the link for the finetuned BB of Stanford-CXR.

(b) Generating dataset for Semi-supervised learning (SSL)

As mentioned in the Algorithm 1, we need to generate the concepts for Stanford-CXR using Semi-supervised learning (SSL), as Stanford-CXR does not have any concept annotations or radiological reports associated to it. Also, to prove MoIE-CXR's efficiency, we use a subset of training data of Stanford-CXR. Refer to the link to get csvs with details of the subset of 5000, 11000, 15000, 20000 and 30000 training samples of Stanford-CXR and place it in the directory Path/out/stanford_cxr/BB/lr_0.01_epochs_5_loss_CE/densenet121/<disease>. For example, to get the details of 15000 training samples of cardiomegaly of Stanford-CXR, place the master_tot_15000.csv file in the directory Path/out/stanford_cxr/BB/lr_0.01_epochs_5_loss_CE/densenet121/cardiomegaly. The lr and the epoch in the path is the learning rate and epoch at which the BB for Stanford-CXR has been trained. You can get it from the 1st command of each script in the folder ./src/scripts/scripts_stanford_cxr.

If you want to generate the csvs with the samples of Stanford-CXR, here is an example command to generate and save the details of 30000 samples from the training data of Stanford-CXR:

python .src/codebase/prepare_df_SSL_main.py \
  --dataset="stanford_cxr" \
  --disease="cardiomegaly" \
  --pos_correct 2200 \
  --neg_correct 11600 \
  --pos_incorrect 15200 \
  --neg_incorrect 1000

While training using SSL, master_tot_15000.csv file is required in the dataloader for the training script in ./src/codebase/train_t_ssl_main.py. Here are the details of the argparse variables:

• --disease: disease in Stanford-CXR.
• --pos_correct: Correctly classified training samples with the disease (y=1) of Stanford-CXR using the finetuned BB from MIMIC-CXR.
• --neg_correct-root: Correctly classified training samples without the disease (y=0) of Stanford-CXR using the finetuned BB from MIMIC-CXR.
• --pos_incorrect: Incorrectly classified training samples with the disease (y=1) of Stanford-CXR using the finetuned BB from MIMIC-CXR.
• --neg_incorrect: Incorrectly classified training samples without the disease (y=0) of Stanford-CXR using the finetuned BB from MIMIC-CXR.

These 4 numbers will add up to --tot_samples in the subsequent scripts in next stage.

(c) Finetuning MoIE-CXR for transfer learning

• Go the ./src/scripts/scripts_stanford_cxr folder and get the training scripts. For all the diseases in for Stanford-CXR, one script is included in the respective folder with proper instructions to run 1) Blackbox 2) projection (t) 3) interpretable model (g) 4) residual (r). Fix the number of samples to be used as training data for Stanford-CXR. For example, to run cardiomegaly for Stanford-CXR, look into the file ./src/scripts/scripts_stanford_cxr/cardiomegaly/car_15000.sh and ./src/scripts/scripts_stanford_cxr/cardiomegaly/car_fl_15000.sh to finetune the model and estimate the computation cost. These two files use 15000 samples of Stanford-CXR for the finetuning. To modify this number, use the following variable --tot_samples in those scripts.
• The naming convention and the paths to be replaced is mentioned in the script. Follow them carefully.
• Run them sequentially.
• Also, for training MoIE-CXR using the Stanford-CXR in the script, ./src/codebase/train_explainer_cxr_domain_transfer.py , the variable --initialize_w_mimic signifies whether we finetune both the selectors (pi) and the experts (g) or not.
  • if --initialize_w_mimic == "y", we finetune both the selectors (pi) and the experts (g).
  • if --initialize_w_mimic == "n", we finetune only the selectors (pi), not the experts (g).

(d) Generate FOLs for the samples of Stanford-CXR

Refer below for generating FOLs for the samples of Stanford-CXR where 15000 training samples were used for finetuning for the disease edema. Modify tot_samples and disease for other configurations.

python ./src/codebase/FOL_cxr_domain_transfer.py --iteration 1 --disease "edema" --model "MoIE" --tot_samples 15000 --cov 0.4 --initialize_w_mimic "y"
python ./src/codebase/FOL_cxr_domain_transfer.py --iteration 2 --disease "edema" --model "MoIE" --tot_samples 15000 --cov 0.3 --initialize_w_mimic "y"
python ./src/codebase/FOL_cxr_domain_transfer.py --iteration 3 --disease "edema" --model "MoIE" --tot_samples 15000 --cov 0.3 --initialize_w_mimic "y"

Suggestions

Most of the argparse variables are self-explanatory. However, in order to perform the experiments successfully, give the correct paths and files to the following variables in train_explainer_<dataset>.py and test_explainer_<dataset>.py.

• For train_explainer_mimic_cxr.py and train_explainer_cxr_domain_transfer.py, follow the rules:

  1. --checkpoint-model : Don't include this variable for the 1st iteration. For 2nd iteration and onwards, include the checkpoint files of all the experts of previous iterations while training for the expert ( g) (--expert-to-train "explainer"). For example: if the current iteration is 3, include the checkpoint files for the expert 1 and expert 2 sequentially. While training the residual (--expert-to-train "residual"), include the checkpoint files of all the experts including the current iteration.
  2. --checkpoint-residual : Don't include this variable for the 1st iteration. For 2nd iteration and onwards, include the checkpoint files of all the residuals of previous iterations while training the expert ( g) (--expert-to-train "explainer") and the residual (--expert-to-train "explainer"). For example: if the current iteration is 3, include the checkpoint files for the residual 1 and residual 2 sequentially.
  3. --prev_chk_pt_explainer_folder : Don't include this variable for the 1st iteration. For 2nd iteration and onwards, include the folders of the checkpoint files of all the experts of previous iterations. For example: if the current iteration is 3, include the checkpoint folders for the expert 1 and expert 2 sequentially. For Include the experiment folder where the checkpoint file will be stored.

• For test_explainer_mimic_cxr.py and test_explainer_cxr_domain_transfer.py, follow the rules:

  1. --checkpoint-model : Don't include this variable for the 1st iteration. For 2nd iteration and onwards, include the checkpoint files of all the experts including the current iteration while testing the expert ( g) (--expert-to-train "explainer") and the residual (--expert-to-train "explainer").
  2. --checkpoint-residual : Don't include this variable for the 1st iteration. For 2nd iteration and onwards, include the checkpoint files of all the residuals of previous iterations while training for the expert ( g) (--expert-to-train "explainer")**. For example: if the current iteration is 3, include the checkpoint files for the residual 1 and residual 2 sequentially. While testing the residual (--expert-to-train "residual"), include the checkpoint files of all the residuals including the current iteration.
  3. --prev_chk_pt_explainer_folder : Don't include this variable for the 1st iteration. For 2nd iteration and onwards, include the folders of the checkpoint files all the experts of previous iterations. For example: if the current iteration is 3, include the checkpoint folders for the expert 1 and expert 2 sequentially. Include the experiment folder where the checkpoint file will be stored.

Also make sure the following variables are correct:

• --cov: Coverages of each iteration separated by a space as in the above commands.
• --lr: Learning rates of each expert separated by a space as in the above commands.
• --data-root: Dataset path of images, labels and concepts (if exists)
• --logs: Path of tensorboard logs

How to Cite

• Main paper MICCAI, 2023

@InProceedings{10.1007/978-3-031-43895-0_59,
author="Ghosh, Shantanu
and Yu, Ke
and Batmanghelich, Kayhan",
editor="Greenspan, Hayit
and Madabhushi, Anant
and Mousavi, Parvin
and Salcudean, Septimiu
and Duncan, James
and Syeda-Mahmood, Tanveer
and Taylor, Russell",
title="Distilling BlackBox to Interpretable Models for Efficient Transfer Learning",
booktitle="Medical Image Computing and Computer Assisted Intervention -- MICCAI 2023",
year="2023",
publisher="Springer Nature Switzerland",
address="Cham",
pages="628--638",
abstract="Building generalizable AI models is one of the primary challenges in the healthcare domain. While radiologists rely on generalizable descriptive rules of abnormality, Neural Network (NN) models suffer even with a slight shift in input distribution (e.g., scanner type). Fine-tuning a model to transfer knowledge from one domain to another requires a significant amount of labeled data in the target domain. In this paper, we develop an interpretable model that can be efficiently fine-tuned to an unseen target domain with minimal computational cost. We assume the interpretable component of NN to be approximately domain-invariant. However, interpretable models typically underperform compared to their Blackbox (BB) variants. We start with a BB in the source domain and distill it into a mixture of shallow interpretable models using human-understandable concepts. As each interpretable model covers a subset of data, a mixture of interpretable models achieves comparable performance as BB. Further, we use the pseudo-labeling technique from semi-supervised learning (SSL) to learn the concept classifier in the target domain, followed by fine-tuning the interpretable models in the target domain. We evaluate our model using a real-life large-scale chest-X-ray (CXR) classification dataset. The code is available at: https://github.com/batmanlab/MICCAI-2023-Route-interpret-repeat-CXRs.",
isbn="978-3-031-43895-0"
}

• Workshop paper published in Interpretable Machine Learning in Healthcare workshop, ICML 2023

@inproceedings{ghosh2023bridging,
  title={Bridging the Gap: From Post Hoc Explanations to Inherently Interpretable Models for Medical Imaging},
  author={Ghosh, Shantanu and Yu, Ke and Arabshahi, Forough and Batmanghelich, Kayhan},
  booktitle={ICML 2023: Workshop on Interpretable Machine Learning in Healthcare},
  year={2023}
}

License and copyright

Licensed under the MIT License

Copyright © Batman Lab, 2023

Contact

For any queries, contact: shawn24@bu.edu