-
Vector Databases - used in AI/ML
-
HIMSS 2023 Feedback
- Weird to be in McCormick Place and it not be RSNA
- LLM / ChatGPT / AI/ML
- Epic/MSFT ChatGPT integration
- Felt very "vendor heavy"
- AI more mature
- More focus on solutions - what is meaningful to implement
- "This is the last year of HIMSS before ChatGPT disruption"
- Lots of vendors products will be replaced by ChatGPT solutions?
- On the cusp of crazy change
- Attendance ~ 40K - less than RSNA at its peak
- More healthcare providers than vendors
- Culture is different - more C level execs attending or presenting
- More receptive to putting high level execs on stage to do talks
- RSNA makes presentations more generic / no branding
- Interoperability showcase - lots of floorspace (4x the size of the AI one at RSNA)
- Hot topics
- Provider fatigue - hungry for any solution to help with this
- Human resource shortage
- ChatGPT Language Translation
- Critcisms:
- Shuttle service sucked
- Frequency was lower
- Missing Carpets!
- Booth numbering was nonsense
- Booths were smaller
- Eisle ways were half - 1/3 as wide
- Shuttle service sucked
-
DICOM Pixel Formats - signed, unsigned, bits allocated, bits stored, high bit, number of components, photometric interpretation, endianess
- How many of these are relevant w.r.t. backwards compatibility
- Do you have to deal with this still? YES
- Implicit LIttle Endian is the only transfer syntax you can count on
- AHLI tries to simlify this and transcodes everything to HTJ2K
- NOTE: Compression algorithm headers override DICOM attributes
- Movement in IHE to go towards FHIR Implementation Guide (IG) development of profiles
- Challenges
- FHIR standard is moving very quickly
- Measuring standards compliance
- IHE Compliance tool "Gazelle"
- Groups trying to turn Gazelle into a FHIR compliance testing tool
- Has not been a lot of collaboration between DICOM and HL7/FHIR recently
- There should be more
- Should be better linking between FHIR and DICOM
- Challenges
- FHIR Implementation conformance framework:
- DICOM Schema is "huge"
- 4000+ attributes, many of which are not required
- Multi-frame vs single frame
- Private attributes
- HL7 FIR r5 ImagingSelection Resource
- Would be nice to have a layer that "normalizes" the pixel data?
- CornerstoneJS does this to an extent because JavaScript is not strongly typed (pixel data is an array of numbers)
- Doing more normalization than this would be probelmatic for some use cases
- CornerstoneJS does this to an extent because JavaScript is not strongly typed (pixel data is an array of numbers)
- https://github.com/PantelisGeorgiadis/dcmjs-imaging