Overview of supported workflows

We describe two separate workflows that share common steps when launching training but differ in the initial configuration, data import and data preparation steps. Describing two workflows implies a binary choice for the user, but you will likely perform some of the steps in a different order, iterate through parts of the process and possibly combine concepts from the workflows. Please read the two workflows and understand that they provide guidance but do not define the only path through an overall process.

The workflows also assume the specific task of segmenting organs. Other tasks will have slightly different requirements.

General Workflow Steps for Organ Segmentation

  1. Import data from PACS
  2. Segment images with consistent labels using OHIF viewer; this constitutes "truth"
  3. Convert images and segmentations into format required for your model training
  4. Start with a pre-trained model and iterate through the training process, tuning the model in the process
  5. Export/publish model results for you or another user to perform further training or to run inference on clinical or research data.

"Pre-Coordinated" Workflow

The Pre-Coordinated Workflow makes the assumptions listed immediately below. When we refer to knowledge or skills that you possess, that can mean either you explicitly or an associate with a specific set of skills. You need to understand the overall process. As an example, you might rely on associate to accurately draw the segmentation outlines of organs.

Assumptions

  • You have identified a set of DICOM studies that exist on a clinical or research PACS. Your XNAT is configured and authorized to perform DICOM C-FIND and DICOM C-MOVE operations, through the DQR plugin or other means.
  • You know the set of organs that you want to segment and have defined text labels that will be used throughout the processing of drawing the regions of interest.
  • You have identified an existing model on the NVIDIA GPU Cloud (NGC) website or from some other source.
  • You understand the input requirements for the model you have selected and understand how to configure XNAT to produce those required inputs.
  • You understand how to adjust the inputs for the Clara Train software and iterate through the process of refining the original model.

Workflow

The Pre-Coordinated workflow follows the steps listed above. Here is a detailed walkthrough: Pre-coordinated Workflow


Retrospective Workflow

Assumptions

  • You have DICOM images in an existing XNAT project that are suitable for the segmentation task.
  • The project already includes RT Struct objects with consistent organ labels produced by the OHIF viewer.
  • You have identified an existing model on the NVIDIA GPU Cloud (NGC) website or from some other source.
  • You understand the input requirements for the model you have selected and understand how to configure XNAT to produce those required inputs.
  • You understand how to adjust the inputs for the Clara Train software and iterate through the process of refining the original model.

Workflow

You can imagine variants in the starting points for your project. For example:

  • You might have an existing XNAT project with images but no RT Struct objects, or
  • You might need to push DICOM data to your XNAT because you are not allowed access to data on a PACS, or
  • You are not starting with a pre-trained model

We trust you will make appropriate adjustments as you proceed. Here is a detailed walkthrough of one such path: Retrospective Workflow

$label.name