XNAT ML Documentation
Deprecation Notice
The XNAT ML Plugin suite was built around the now-deprecated NVIDIA Clara 2.0 style of models and training. These plugins have been deprecated and are not recommended for production use but are still available as open source resources for those who are interested.
The XNAT ML and Datasets plugins were the result of more than a year’s intensive development that began with a demo challenge partnering with NVIDIA for RSNA 2019. The beta release followed in June 2020, and the full production version of the ML plugin set was released in concert with XNAT 1.8.0 in March 2021. This work was the XNAT team’s first foray into actively supporting machine learning workflows, which are a critical path for advancing research into predictive medicine. So it is with mixed feelings that we announce the deprecation of this plugin suite.
Why Deprecate?
The ML plugin architecture was built around the NVIDIA Clara 2.0 set of models, supporting their training and configuration paradigm. However, nearly as soon as we released the production 1.0 version, NVIDIA deprecated these models in favor of version 3.0, which soon fell by the wayside themselves in favor of even more advanced model frameworks. It soon became clear that our pace of work could not keep up with the NVIDIA engineering team. We maintained support for these plugins as proofs of concept, but ultimately their utility has waned.
In the meantime, XNAT community development into the ML world has taken different forms:
The XNAT-OHIF Viewer plugin, developed at ICR UK, introduced support for NVIDIA’s MONAI Label suite of AI-Assisted Annotation tools
The XNAT JupyterHub Integration plugin, developed at Washington University in St Louis, offers expanding support for mounting XNAT data into a Jupyter Notebook and performing exploratory model development
How To Remove the ML and Datasets Plugins
Since both the ML and Datasets plugins introduce new datatypes to the XNAT schema, removing the plugins is not as simple as simply deleting the jar files and restarting tomcat. Once datatypes are installed and new Postgres tables are created, XNAT gets cranky if the XML Schema Document (XSD) is removed.
The simple way to address this is to create and install schema-only plugins to install in place of the originals. These plugins will sit silently in your file system, preserving the connection between XNAT and the legacy datatype.
Once you replace these plugins with schema-only alternatives, the following effects will take place:
The APIs for ML and Datasets will be removed
Uploading of new models and training configurations will be disabled
Creation of new dataset criteria and collections in a project will be disabled
Model training will be disabled
The ML and Datasets datatypes will still be registered in Datatype Administration, however search listings for these datatypes will no longer be active
