2022-02-28.md

2022-02-28

Agenda

• SIIM Hackathon Plug (Mohannad)

  • May 31-June 3
  • Focused on education and collaboration above all else (vs competition)
  • Big fan of open source - use open source products and open source data sets
  • New and interesting things this year
    • Bring in non radiology images - pathology, dermatology, opthtalmology|ophthalmology?
    • Blockchain demystified
  • OK to team up/pair with others at the
  • OK to just hang around and observe
  • Slack Channel: https://join.slack.com/t/siimhackathon/shared_invite/zt-mkk0yn2e-KUqOLi6ETBUQmOffxmcQxA
  • Particpation is 100% free - but you need an API key (free)
  • Various virtual meetings dureing the hackathon
    • Orientation (2 weeks before hand)
    • Brainstorm meeting (May 31)
    • Check in meetings 2x daily

• Segmentation (automatic, semi-automatic, manual)

  • Good open source projects (automatic and semi-automatic)
    • https://github.com/Project-MONAI/MONAILabel
      • Nice tool to put your monai labels from nVidia/Kitware
      • Allows you to deploy the models in 3D Slicer and OHIF
      • Nice feature for continuous learning (make corrections and send them back to model to train it/make it better)
      • Built on top of pytorch
      • Several models and architectures available on github
        • Some examples for https://github.com/Project-MONAI/tutorials segmentation architectures
    • Microsoft Inner Eye
      • https://www.microsoft.com/en-us/research/project/project-innereye-open-source-software-for-medical-imaging-ai/
    • OHIF
      • basic manual segmentation tools
    • 3DSlicer
      • ??
    • https://medseg.ai
      • https://www.linkedin.com/in/tomas-sakinis/
      • All browser side (tensorflow.js?)
    • VGG's VIA Tool (standalone html, manual)
      • https://www.robots.ox.ac.uk/~vgg/software/via/
    • https://www.crowds-cure.org/
  • Good commercial offerings
    • md.ai
    • Axial3D
      • Taking DICOMs - run inference against this, produces segmentation
      • Have a whole team of biomed engineers that adjust the segmentation as needed
      • Generate meshes from the segmentation - can be rendered in web app or sent to 3D printer
      • Have a set of models trained against labels generated by biomed engineers over the years
      • Use Cases:
        • Educational
        • Moving towards planning for surgerys
        • Discovering new ones
          • Hospital networks..
  • DICOM SEG
    • Any experience generating or reading these?
    • Andre Fidorav has a toolkit for this - import/export between Nifti and DICOM SEG. Also has DICOM-SR TID1500 and parametric maps (for quantification)
      • https://github.com/QIICR/dcmqi

• SIIM - June 9-11 in person also virtual option?

  • https://siim.org/page/meetings

• Is there a platform for moderately technical people in terms of model training to do less sophisticated model training for classification (e.g. how to tell MR from CT)

  • Google AutoML, AWS SageMaker
  • Once you train the model, can you export this into tensorflow.js so you can run in browser?
    • Google was locked into their service (at least of a few years ago)
  • most effort goes into data pipeline - normalization, anonymization
  • In one of RSNA 2021 course this was used https://teachablemachine.withgoogle.com/train/image