Skip to content

Latest commit

 

History

History
49 lines (42 loc) · 4.25 KB

2022-03-28.md

File metadata and controls

49 lines (42 loc) · 4.25 KB
Raw

2022-03-28

Agenda

  • Question from Adam: Processing and anonymizing DICOM, segmentation. Recently added support for multiframes - splitting them into single DICOMs. What is the standard way to deal with multiframes?
    • Simon: multiframe didn't catch-on as we would have liked. Seen most ML implementations split them up. There are libraries to do that, e.g. dcm4che (converting both ways multiframe <-> single frame)
    • Mohannad pointed out difference between dcm4chee vs dcm4che, as well as 2.x vs 5.x branches (entirely different, total re-write)
    • Neil pointed out a Twitter community for ML in Healthcare: https://twitter.com/i/communities/1506797413922750466. Pretty new feature of Twitter. Not a fan of Twitter but trying it out. This community is moderated (currently the only moderator but happy to expand).
    • Jean-Francois: could also transcode to something that's really fast to load, and track the offsets of the frames.
    • Neil: Transformations and 2D vs 3D models could change you'd process a series of images.
    • Sandor: In recent/older Kaggle contests pretty good results were achieved by 2.5D models. Changing a 2D model with a (R,G,B) channel image as input could be changed to use a greyscale (layer-1, layer, layer+1) structure. SO it got kind-of "volumetric aware"
    • Mohannad: Always tricky with communities, e.g. Slack or Discord vs. Mailing lists (aren't we all drowning in emails?)
    • Simon: A number of communities actively use Discord. Link to this comunity's Discord: https://discord.gg/GZmcTRT3
  • Interactive Multimedia Reportings (IHE_IMR profile)
  • Link to profile: https://profiles.ihe.net/RAD/IMR/index.html.
  • Published a couple of weeks ago for public comments. Go to IHE Radiology and you'll see AIR+ and this profile, IMR. Comments needed by September, in time for the IHE connectathon.
  • FHIR DiagnosticReport as the framework to create a new report. DICOMweb first, DIMSE second for image references.
  • Providing a way for report creators, EMRs and PACS to integrate
  • Also builds on top of IID (Invoke Image Display)
  • Have buy-in for Nuance and MModal, hoping to demonstrate at RSNA (Imaging AI in Practice demo). Other smaller vendors too.
  • Will present more info at a later date.
  • Kinson Ho and Seth Berkowitz are lead author
  • Three scenarios for integrations, e.g.:
    • Highly integrated, interactive viewer
    • In the EMR, with a zero-footprint viewer
  • Invoker actor, facilitates communication between PACS and display
  • Fallback to OBX segment
  • Mohannad: I think IID is due for overhall. It is good enough for IMR, but beyond that (general workflows), it is due for updates. IID came out 7+ years, when a URL integration was cutting edge, but nowadays we should be able to do a lot more with things like FHIRcast
  • David: Kinson thought about that, but decided against it
  • Neil: What's the primary usecase? E.g. patient views their report and we want to show them images in-line?
  • David: Definitely a usecase but plenty of others. E.g. referncing a prior study/report. Also the referring doctor/GP viewing the report. Were very ambitious but had to scale it back for a number of issues, security being a prominent one.
  • Simon: Incorporating FHIR ImagingSelection?
  • Seth: Yes, this is a first step (putting a flag in the ground), future iterations would grow to include features like this.
  • David: PDF vs. interactive HTML document. Trying not to leave any legacy systems behind. Multiple formats supported.
  • Mohannad: Europe was way ahead as far as patient portal implementations, should have helpful input here.