/mDKL

The PyTorch code for the publication "Quantifying Predictive Uncertainty in Medical Image Analysis with Deep Kernel Learning" on ICHI 2021 (9th IEEE International Conference on Healthcare Informatics)

Primary LanguagePythonMIT LicenseMIT

Description

The Python implementation for the publication “Quantifying Predictive Uncertainty in Medical Image Analysis with Deep Kernel Learning” on ICHI2021 (the 9th IEEE International Conference on Healthcare Informatics) using PyTorch and GPyTorch packages.

Project structure

.
├── cae_ba.py   
├── cae_dl.py    
├── cae_utils.py
├── data_utils.py
├── densenet.py    # a modified densenet to enable dropout behaviour during the inference phase
├── gp_layer.py    # SVGP-based output prediction layers
├── LICENSE
├── logging_conf.py
├── model_utils.py
├── plot_utils.py
├── pml_ba.py    
├── pml_dl.py  
├── requirements.txt
├── run_exp_ba_cae.py
├── run_exp_ba_dropout.py
├── run_exp_ba_metric.py
├── run_exp_ba.py
├── run_exp_dl_cae.py
├── run_exp_dl_dropout.py
├── run_exp_dl_metric.py
├── run_exp_dl.py
├── run_exp_gp_ba_cae.py
├── run_exp_gp_ba_metric.py
├── run_exp_gp_ba.py
├── run_exp_gp_dl_all_cae.py
├── run_exp_gp_dl_all_metric.py
└── run_exp_gp_dl_all.py

Usage

  • All .py files should be able to run with python xxx.py after installing the packages specified in requirements.txt.
  • The .py scripts prefixed with run_exp_ can be used to generate (similar) results in Table I/II/III/IV.
    • Dataset has to be downloaded for the RSNA Bone Age dataset and the Deep Lesion dataset.
    • Scripts with …_ba_… are for the experiments with the RSNA Bone Age dataset.
    • Scripts with …_dl_… are for the experiments with the Deep Lesion dataset.

Note

The code is published to ensure the reproducibility in the machine learning community. If you find the code helpful, please consider citing

@article{wu2021quantifying,
  title={Quantifying Predictive Uncertainty in Medical Image Analysis with Deep Kernel Learning},
  author={Wu, Zhiliang and Yang, Yinchong and Gu, Jindong and Tresp, Volker},
  journal={arXiv preprint arXiv:2106.00638},
  year={2021}

License

The code has a MIT license, as found in the LICENSE file.