Generating and Evaluating Post-Hoc Explanation from Deep Neural Networks for Multi-Modal Medical Image Analysis Tasks
This is the software repository for our work that evaluates 16 commonly-used explainable AI algorithms on a novel problem of multi-modal medical image analysis tasks.
We generate explanations on two multi-modal medical image analysis tasks: brain glioma grading task on the BraTS brain MRI dataset, and knee lesion classification on the MRNet knee MRI dataset.
├── code
│ ├── xai # The main code to generate and evaluate explanations using different post-hoc heatmap algorithms
│ ├── data_loader # Dataloader for the BraTS data (data_loaders.py) and the synthetic glioma data (tumorgenerator_dataloader.py)
│ ├── model # GeneNet is the VGG-like 3D model on the brain glioma MRI BraTS data
│ ├── TumorSim # Utility function for the tumorgenerator_dataloader.py of the synthetic glioma data. Code from http://dx.doi.org/10.1002/mp.14701
│ ├── MRNet # Model training and testing for the MRNet dataset on knee lesion classification
│ ├── utils # Utility function for model training
│ ├── sh # Bash and json files for experiment running
│ ├── xai_pipeline.py # Heatmap explanation generation and evaluation pipeline for the BraTS model
│ └── xai_mrnet.py # Heatmap explanation generation and evaluation pipeline for the MRNet model
├── data # Train, validation, and test data split csv files for MICCAI_BraTS2020_TrainingData
├── paper # Preprint paper and supplementary materials
├── image # Graphic abstract of the associated papers
├── neurosurgeon_clinical_study_data_analysis # Data analysis code for the paper "Evaluating the Clinical Utility of Artificial Intelligence Assistance and its Explanation on the Glioma Grading Task"
└── README.md
- Providing explanation is important for the clinical deployment of AI-based decision support systems
- We conduct computational evaluations to examine whether the existing post-hoc explainable AI algorithms can fulfill clinical requirements
- Generating explanation for multi-modal medical image tasks
- Evaluating explainable AI algorithms based on clinical requirements as outlined in the Clinical Explainable AI Guidelines
- Evaluating the clinical utility of AI and explanation with 35 neurosurgeons on the glioma grading task.
- Code for study data analysis
- The clinical study data neurosurgeon35_data.csv is undergoing ethics review. We will make the study data publicly available once the ethics application is approved.
git clone [email protected]:weinajin/multimodal_explanation.git
cd multimodal_explanation
# install the requirements
conda create -n brain python=3.7
conda activate brain
conda install -c pytorch torchvison=0.5
pip install -r ./code/requirement_cc.txtTraining the model requires downloading the publicly-available BraTS 2020 dataset for the glioma grading task, or MRNet dataset for the knee lesion classification task.
- Training model on the BraTS data
python train.py --config sh/config_cc_plain2_BRATS_HGG.json --fold 1 --seed 2
- Training model on the synthetic glioma data
python train.py --config sh/tumorsyn_solar.json --seed 2
- Training model on the knee MRI data
python train.py --task meniscus --seed 2
Generating and evaluating the heatmaps require to have the trainde model. In addition, the informative plausibility evaluation will need the BraTS 2020 dataset dataset, which contains the ground-truth tumor segmentation mask.
The --job parameter can be spcified to run for different jobs listed below:
gethm: Generate heatmap explanation for the explainable AI algorithm as specified in the json file:xai/method_list. The generated heatmaps are saved as .pkl files for further evaluation.mi: Run modality ablation experiments to get the modality Shapley value as modality importancemi_reporting: Compute the Shapley value for each medical image modality as modality importancemi_readhm: Calculate the heatmap sum for each image modalityfp_iou: Calculate the feature portion (fp) and intersection over union (iou) by comparing the heatmap with ground-truth feature segmentation masksmsfi_readhm: Calculate the feature portion for each image modality for the MSFI metricacc_drop: Run the cumulative feature removal experimentpipeline: Run the above functions as a pipeline to generate and evaluate heatmapspipeline_nogethm: Run the above heatmap evaluation pipeline without generating heatmapsmi_corr: Calculate the modality importance correlation
Example:
python xai_pipeline.py --config sh/xai_cc_plain2_BRATS_HGG.json --fold 1 --seed $SEED --bs 1 --job <job>Guidelines and evaluation of clinical explainable AI in medical image analysis
Weina Jin, Xiaoxiao Li, Mostafa Fatehi, Ghassan Hamarneh. Medical Image Analysis. 2023
This paper include the two evaluations on the brain and knee datasets.
@article{JIN2022102684,
title = {Guidelines and evaluation of clinical explainable AI in medical image analysis},
author = {Weina Jin and Xiaoxiao Li and Mostafa Fatehi and Ghassan Hamarneh},
journal = {Medical Image Analysis},
volume = {84},
pages = {102684},
year = {2023},
issn = {1361-8415},
doi = {https://doi.org/10.1016/j.media.2022.102684},
url = {https://www.sciencedirect.com/science/article/pii/S1361841522003127},
}
Weina Jin, Xiaoxiao Li, Ghassan Hamarneh. AAAI. 2022
This paper is the evaluation on the brain dataset alone.
@article{Jin_Li_Hamarneh_2022,
title = {Evaluating Explainable AI on a Multi-Modal Medical Imaging Task: Can Existing Algorithms Fulfill Clinical Requirements?},
author = {Jin, Weina and Li, Xiaoxiao and Hamarneh, Ghassan},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
url = {https://ojs.aaai.org/index.php/AAAI/article/view/21452},
DOI = {10.1609/aaai.v36i11.21452},
year = {2022}, month = {Jun.}, number = {11}, volume = {36},
pages = {11945-11953}
}
Weina Jin, Xiaoxiao Li, Mostafa Fatehi, Ghassan Hamarneh. MethodsX. 2023
This paper describes the review and implementation of the included post-hoc explanation algorithms.
@article{JIN2023102009,
title = {Generating post-hoc explanation from deep neural networks for multi-modal medical image analysis tasks},
journal = {MethodsX},
volume = {10},
pages = {102009},
year = {2023},
issn = {2215-0161},
doi = {https://doi.org/10.1016/j.mex.2023.102009},
url = {https://www.sciencedirect.com/science/article/pii/S2215016123000146},
author = {Weina Jin and Xiaoxiao Li and Mostafa Fatehi and Ghassan Hamarneh},
}
Weina Jin (co-first author), Mostafa Fatehi (co-first author), Ru Guo, Ghassan Hamarneh. Artificial Intelligence in Medicine. 2024
This paper is the neurosurgeon clinical user study to evaluate the clinical utility of AI and explainable AI on task performance. Video paper presentation
@article{JIN2024102751,
title = {Evaluating the clinical utility of artificial intelligence assistance and its explanation on the glioma grading task},
journal = {Artificial Intelligence in Medicine},
pages = {102751},
year = {2024},
issn = {0933-3657},
doi = {https://doi.org/10.1016/j.artmed.2023.102751},
url = {https://www.sciencedirect.com/science/article/pii/S0933365723002658},
author = {Weina Jin and Mostafa Fatehi and Ru Guo and Ghassan Hamarneh},
keywords = {Artificial intelligence, Neuro-imaging, Neurosurgery, Explainable artificial intelligence, Clinical study, Human-centered artificial intelligence},
abstract = {Clinical evaluation evidence and model explainability are key gatekeepers to ensure the safe, accountable, and effective use of artificial intelligence (AI) in clinical settings. We conducted a clinical user-centered evaluation with 35 neurosurgeons to assess the utility of AI assistance and its explanation on the glioma grading task. Each participant read 25 brain MRI scans of patients with gliomas, and gave their judgment on the glioma grading without and with the assistance of AI prediction and explanation. The AI model was trained on the BraTS dataset with 88.0% accuracy. The AI explanation was generated using the explainable AI algorithm of SmoothGrad, which is selected from 16 algorithms based on the criterion of being truthful to the AI decision process. Results showed that compared to the average accuracy of 82.5±8.7% when physicians performed the task alone, physicians’ task performance increased to 87.7±7.3% with statistical significance (p-value = 0.002) when assisted by AI prediction, and remained at almost the same level of 88.5±7.0% (p-value = 0.35) with the additional assistance of AI explanation. Based on quantitative and qualitative results, the observed improvement in physicians’ task performance assisted by AI prediction was mainly because physicians’ decision patterns converged to be similar to AI, as physicians only switched their decisions when disagreeing with AI. The insignificant change in physicians’ performance with the additional assistance of AI explanation was because the AI explanations did not provide explicit reasons, contexts, or descriptions of clinical features to help doctors discern potentially incorrect AI predictions. The evaluation showed the clinical utility of AI to assist physicians on the glioma grading task, and identified the limitations and clinical usage gaps of existing explainable AI techniques for future improvement.}
}
Please create a new issue detailing concisely, yet complete what issue you encountered, in a reproducible way.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent