MHub Runner is a 3D Slicer plugin that seamlessly integrates Deep Learning models from the Medical Hub repository (mhub) into 3D Slicer.
MHub is a repository for Machine Learning models for medical imaging. The goal of mhub is to make these models universally accessible by containerizing the entire model pipeline and standardizing the I/O interface.
Find out more at mhub.ai and check out the mhub GitHub repository.
We only need Docker 🐳 to be installed on your system. That's it.
This plugin is under active development. To test it in 3D slicer follow these steps:
- Clone the repository to a local folder.
- Enable developer mode in 3D Slicer (3D Slicer > Settings > Developer > Enable developer mode).
- Start the Extension Wizzard Module (3D Slicer > Modules > Developer Tools > Extension Wizard)
- In the Extension Wizzard click on Select Extension.
- Navigate to the SlicerMHubRunner folder (you cloned it in step 1) and press open.
- The plugin is setup and the MRunner module is now available in Slicer. To open it, navigate to 3D Slicer > Modules > Examples > MRunner.
First, open a volume in Slicer on which you want to run the plugin. You can use your own data or the slicer sample data. Slicer sample data can be found at File > Download Sample Data, to download a chest CT scan click on CTChest. Now open the MRunner module (navigate to 3D Slicer > Modules > Examples > MRunner). You will now see the graphical user interface (GUI) of the module.
The MRunner module has four sections. The model selector is the first section at the top of the GUI. Here you select one of the models from MHub that you want to run. The second section is the input selection where you can select an input volume to load into the model. The third section is the output selection. Since we are only providing segmentation models for now, you select here where the generated segmentations will be stored. You have several options, for example you can create a new segmentation node or select an existing one which will be overwritten. The fourth section offers some advanced options. Here may can check the Use GPU checkbox to download a cuda-enabled image and run the model with GPU acceleration.
Download Dockerfile
Dockerfiles are pulled from our Dockerhub repository (mhubai) via docker pull. Instead, you can build the image locally on your machine by downloading the dockerfile into a temporary directory.
Use GPU
For all models that support GPU acceleration, enable the UseGPU flag to run the model in GPU mode. This flag is set by default for supported models. When it is set, we use a Docker base image that bundles all requirements, e.g., cuda. If you are using a machine without a GPU, you can uncheck this option to use the lightweight base image instead. Note that not all models support CPU-only mode. For those models that do, the execution time is likely to be much higher when running without GPU acceleration.
Docker build --no-cache
During execution, we always check if the image is already available on the local system. If it is not, we initiate either a Docker pull or a Docker build after downloading the Dockerfile. If you select this option, a push or build will be forced even if the image is already available locally. If the Download Dockerfile option is enabled, the image is downloaded and built locally with the --no-cache flag set.
This repository and plugin are under active development, as is the mhub repository. Use this plugin with caution and always backup your data. We strongly recommend that you only use the slicer sample data and only use this plugin in a non-production environment.