2022-02-7.md

2022-02-7

Agenda

• Nvidia

  • Monai (open source AI Libraries)
    • Training AI models for medical imaging and beyond
    • Labeling component that plugs into OHIF
      • Active Learning
    • Deployment Framework - stream together a model into a full application
      • What is the cope for this - is it intended to be a comprehensive DICOM Ingestion processing framework?
        • NO - may be part of this though
        • https://docs.monai.io/projects/monai-deploy-app-sdk/en/latest/release_notes/index.html
        • https://docs.monai.io/projects/monai-deploy-app-sdk/en/latest/introduction/architecture.html
        • Inference service - https://github.com/Project-MONAI/monai-deploy-app-server/blob/main/components/inference-service/README.md
        • NOT intended to be single endpoint that runs lots of algorithms OR do metadata fixup (e.g. uncompress images)
          • Does not take the place of dicom routers or tag morphing
        • Focused on running jobs effectively that is configured
      • How does this relate to the DICOM Standard for AI/ML processing (WG23)
        • Team is playing close attention to this
        • Current work item on discovery and control - in flux so uncertain where it will end up
        • Looking at IHE profiles (AIWI and AIR)
        • Looking at work being done in standards and incorporating it in a practical way
        • Lessons learned are being fed back in to the standards organizations
  • FLAIR (Open source federated learning)
    • Framework goes beyond medical imaging (could be applied to anything)
    • Allows for running jobs collectively from different sites
    • Want data to stay at the organization that owns it - moving compute to data
  • Clara Viz (not open source)
    • Framework for rendering medical images in 3D space
    • SDK
    • Clara Viz - https://github.com/NVIDIA/clara-viz
  • All components are super easy to integrate with your language (e.g. pip install for python)
  • GTC coming up next month (twice a year)
    • Virtual (free)
    • Medical Imaging specific sessions
      • Bootcamp on Monai, etc

• Visualization of volumes in time

  • Knows OHIF is doing something here (Cornerstone3D)
  • Is there anything out there?
    • Slicer 3D extension - ideal for pediatric data sets?
      • Echocardiograms can be DICOM but also some other "funky" format

• 7 Tesla MRI of the ex vivo human brain at 100 micron resolution

• Youtube video: https://www.youtube.com/watch?v=oHZBFm02wbM

• DICOM? DataSet: https://www.pacsbin.com/c/W1IVmnkSAE

• DataSet: https://datadryad.org/stash/dataset/doi:10.5061/dryad.119f80q

• nVidia library for JPEG2000

  • https://developer.nvidia.com/nvjpeg

• Very old cornerstone server side rendering library:

  • https://github.com/chafey/cornerstoneVisualization

• 3D visualization library

  • https://github.com/rapidsai/cucim

• Slicer3D as a "web app"

  • Uses AWS AppStream
  • Microsoft has Windows 365 / Azure VIrtual Desktop which is similar. Has persistent option
  • YouTube Video: https://www.youtube.com/watch?v=u3t5twSV6NE&t=33s
  • https://discourse.slicer.org/t/running-slicer-in-a-web-browser-via-amazon-appstreams/21431

• Lossy compression vs lossless for diagnostic interpretation

  • Simon has an outstanding "challenge" for anyone to show how lossy compression or non diagnostic monitor resulted in misdiagnosis
  • Rads are rarely aware that an image is being displayed lossy (despite indicators on screen)
  • Some vendors are doing lossy display "by default" to hit the performance targets over lower bandwidth/higher latency
  • There was a case where a bug in a PACS system resulted in a difference of about ~8 hounsfield that would have resulted in misdiagnosis had the radiologist missed it
  • is there every a problem with lossy compression that shows things that aren't there (noise/specs/?)
    • yes - happens all the time, even w/o lossy compression