Back to Projects List
- Leonard Nürnberg (Department of Radiology, Brigham and Women’s Hospital, Boston, MA)
- Dennis Bontempi (Department of Radiology, Brigham and Women’s Hospital, Boston, MA)
- Andrey Fedorov (Department of Radiology, Brigham and Women’s Hospital, Boston, MA)
We are working on a repository to standardize Deep Learning models in medical imaging and make them easily accessible to everyone.
A central point of our efforts is to develop a standardized I/O framework for all models to unify the data stream into and out of these models, making them interchangeable. Seamless integration of our repository with Slicer would allow immediate application and exploration of models without the need to set up model environments locally.
Therefore, we are planning a Slicer extension that will allow to search our repository from within Slicer to deploy and run these models locally, using the Slicer interface for data input and output.
Link to the Plugin.
- Objective A. Discover how to run models in Slicer securely, conflict-free, and platform-independent.
- Objective B. Validate and customise our definitions of a generic I/O framework.
- Objective C. Document pros and cons of docker vs native python integration of the model, support with experimental results. Concerns re Docker communicated earlier:
- Docker may be challenging to install and setup (org constraints, permissions, expertise)
- Docker images are large and slow to download
- Support of GPU with Docker is not straightforward/limited
Approach and Plan
- We have two approaches in mind: packaging models in Docker containers and running models in separate Python environments. For both options, we need to weigh the pros and cons to find the most appropriate solution that maximizes the user base while avoiding or minimizing manual per-model customization.
- We want to create a clear understanding of the challenges and drawbacks of using Docker. We will provide a step-by-step guide to setting up Docker and are looking for volunteers to try out the setup under our guidance and report back their valuable feedback.
- We would like to discuss and find the best standard for model outputs (e.g., segmentation label names).
- We plan to develop a slicer plugin that connects data loaded into Slicer to a model via our i/o framework and transfers the model outputs back into Slicer, using selected DL models as proofs of concept.
Progress and Next Steps
- We have developed an experimental modular conversion framework to bridge between a standardized I/O and specific model requirements.
- We dockerized two models (Totalsegmentator, Platipy) using our I/O FW.
Background and References