- About XNAT
- News & Events
- XNAT Marketplace
- Contact Us
The image data, model files and configuration data needed to run training are managed at the XNAT Project level. You will need to create a project in XNAT and consider other users you want to add in the Member or Collaborator role. The standard security considerations are documented here: Understanding Data Sharing in XNAT's Security Structure.
See: Creating a Project and Assigning Roles
The Clara Train software can start with a pre-trained model and supports transfer learning.
See: Installing a Model
The labeling of different types of regions within a volume is typically done with "mask" images, images whose pixel values are integers uniquely assigned to the different regions of interest.
See: Configuring a Segmentation Mapping Resource for your Model
XNAT can be configured to automatically run a workflow that generates resources needed by a model, triggered by the importing of ROI contours into a session. This workflow can only be configured on a single model.
See: Set up an Automated Workflow to Generate Required Model Resources
The Clara Train software operates on files that are in NIFTI format. Medical images are generated by acquisition modalities using the DICOM standard and are uploaded into XNAT through a variety of mechanisms that are described on this page: Image Session Upload Methods in XNAT.
The OHIF viewer, integrated with XNAT, is used to segment and label regions in the image data. The Viewer is used to draw contours on image slices. The contours are imported to XNAT and are converted into the NIFTI-formatted segmentation images. Comprehensive documentation for the Viewer can be found at OHIF-Viewer-XNAT-plugin.
See: Drawing Contours with the XNAT OHIF Viewer
Exporting an ROI Contour Collection to XNAT will trigger automated workflow to generate resources needed by the configured model. This was accomplished by creating event subscriptions that trigger processing by the Container Service. See "Configure Automatic Workflow to Generate Required Model Resources".
Once you have data uploaded and annotated, you need to build a map of the specific XNAT resource files that will be used in training. Step one in this process is defining the model- and project-specific parameters that will build that mapping.
See: Defining Parameters for your Dataset
Step two is validating your project data against that definition, and creating the final dataset.
See: Validating and Saving your Dataset from Project Data
NVIDIA Models (and others) use a JSON configuration that allows you to bake in certain parameters, and allow other parameters to be set at runtime. The XNAT training config goes a step further, and allows you to define other components of stitching your model and dataset together.
See: Creating a Training Config for your Model
With all your prerequisites in place, now you can launch one or more training runs and view their progress in real time.
See: Launching Finetune Training
See: Viewing Training Progress in TensorBoard
With one or more training runs completed, you can review the outputs and determine whether to promote any of those generated checkpoints to become a new model for finetune training.
See: Accessing Training Results
See: Promoting Training Results as a new "Model"
As a final step, you can download your finalized training results for export to a Model Hub, or for use in Inference.