This is a deep learning toolbox to train models on medical images (or more generally, 3D images). It integrates seamlessly with cloud computing in Azure.
On the modelling side, this toolbox supports
- Segmentation models
- Classification and regression models
- Sequence models
- Adding cloud support to any PyTorch Lightning model, via a bring-your-own-model setup
- Active label cleaning and noise robust learning toolbox (stand-alone folder)
Classification, regression, and sequence models can be built with only images as inputs, or a combination of images and non-imaging data as input. This supports typical use cases on medical data where measurements, biomarkers, or patient characteristics are often available in addition to images.
On the user side, this toolbox focusses on enabling machine learning teams to achieve more. It is cloud-first, and relies on Azure Machine Learning Services (AzureML) for execution, bookkeeping, and visualization. Taken together, this gives:
- Traceability: AzureML keeps a full record of all experiments that were executed, including a snapshot of the code. Tags are added to the experiments automatically, that can later help filter and find old experiments.
- Transparency: All team members have access to each other's experiments and results.
- Reproducibility: Two model training runs using the same code and data will result in exactly the same metrics. All sources of randomness like multithreading are controlled for.
- Cost reduction: Using AzureML, all compute (virtual machines, VMs) is requested at the time of starting the training job, and freed up at the end. Idle VMs will not incur costs. In addition, Azure low priority nodes can be used to further reduce costs (up to 80% cheaper).
- Scale out: Large numbers of VMs can be requested easily to cope with a burst in jobs.
Despite the cloud focus, all training and model testing works just as well on local compute, which is important for model prototyping, debugging, and in cases where the cloud can't be used. In particular, if you already have GPU machines available, you will be able to utilize them with the InnerEye toolbox.
In addition, our toolbox supports:
- Cross-validation using AzureML's built-in support, where the models for individual folds are trained in parallel. This is particularly important for the long-running training jobs often seen with medical images.
- Hyperparameter tuning using Hyperdrive.
- Building ensemble models.
- Easy creation of new models via a configuration-based approach, and inheritance from an existing architecture.
Once training in AzureML is done, the models can be deployed from within AzureML or via Azure Stack Hub.
We recommend using our toolbox with Linux or with the Windows Subsystem for Linux (WSL2). Much of the core functionality works fine on Windows, but PyTorch's full feature set is only available on Linux. Read more about WSL here.
Clone the repository into a subfolder of the current directory:
git clone --recursive https://github.com/microsoft/InnerEye-DeepLearning
cd InnerEye-DeepLearning
git lfs install
git lfs pull
After that, you need to set up your Python environment:
- Install
conda
orminiconda
for your operating system. - Create a Conda environment from the
environment.yml
file in the repository root, and activate it:
conda env create --file environment.yml
conda activate InnerEye
- If environment creation fails with odd error messages on a Windows machine, please continue here.
Now try to run the HelloWorld segmentation model - that's a very simple model that will train for 2 epochs on any
machine, no GPU required. You need to set the PYTHONPATH
environment variable to point to the repository root first.
Assuming that your current directory is the repository root folder, on Linux bash
that is:
export PYTHONPATH=`pwd`
python InnerEye/ML/runner.py --model=HelloWorld
(Note the "backtick" around the pwd
command, this is not a standard single quote!)
On Windows:
set PYTHONPATH=%cd%
python InnerEye/ML/runner.py --model=HelloWorld
If that works: Congratulations! You have successfully built your first model using the InnerEye toolbox.
If it fails, please check the troubleshooting page on the Wiki.
Further detailed instructions, including setup in Azure, are here:
- Setting up your environment
- Training a Hello World segmentation model
- Setting up Azure Machine Learning
- Creating a dataset
- Building models in Azure ML
- Sample Segmentation and Classification tasks
- Debugging and monitoring models
- Model diagnostics
- Move a model to a different workspace
- Working with FastMRI models
- Active label cleaning and noise robust learning toolbox
We offer a companion set of open-sourced tools that help to integrate trained CT segmentation models with clinical software systems:
- The InnerEye-Gateway is a Windows service running in a DICOM network, that can route anonymized DICOM images to an inference service.
- The InnerEye-Inference component offers a REST API that integrates with the InnnEye-Gateway, to run inference on InnerEye-DeepLearning models.
Details can be found here.
You are responsible for the performance, the necessary testing, and if needed any regulatory clearance for any of the models produced by this toolbox.
If you have any feature requests, or find issues in the code, please create an issue on GitHub.
Please send an email to [email protected] if you would like further information about this project.
Oktay O., Nanavati J., Schwaighofer A., Carter D., Bristow M., Tanno R., Jena R., Barnett G., Noble D., Rimmer Y., Glocker B., O’Hara K., Bishop C., Alvarez-Valle J., Nori A.: Evaluation of Deep Learning to Augment Image-Guided Radiotherapy for Head and Neck and Prostate Cancers. JAMA Netw Open. 2020;3(11):e2027426. doi:10.1001/jamanetworkopen.2020.27426
Bannur S., Oktay O., Bernhardt M, Schwaighofer A., Jena R., Nushi B., Wadhwani S., Nori A., Natarajan K., Ashraf S., Alvarez-Valle J., Castro D. C.: Hierarchical Analysis of Visual COVID-19 Features from Chest Radiographs. ICML 2021 Workshop on Interpretable Machine Learning in Healthcare. https://arxiv.org/abs/2107.06618
Bernhardt M., Castro D. C., Tanno R., Schwaighofer A., Tezcan K. C., Monteiro M., Bannur S., Lungren M., Nori S., Glocker B., Alvarez-Valle J., Oktay. O: Active label cleaning: Improving dataset quality under resource constraints. ArXiv pre-print (under peer review). https://arxiv.org/abs/2109.00574. Accompagnying code InnerEye-DataQuality
This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.
When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.
This toolbox is maintained by the Microsoft InnerEye team, and has received valuable contributions from a number of people outside our team. We would like to thank in particular our interns, Yao Quin, Zoe Landgraf, Padmaja Jonnalagedda, Mathias Perslev, as well as the AI Residents Patricia Gillespie and Guilherme Ilunga.