-
nVidia GTC - Kicked off today!
- March 20-23
- https://www.nvidia.com/gtc/?ncid=GTC-NV0MZR1P
- Free!
-
Generative AI and Medical Imaging
- Generating Synthetic Image Generation (e.g. "CT Study with liver cancer")
- Registration (e.g 2,000 slice .5mm CT registered with 100 slice 10mm CT)
- Take non contrast CT series and generate a synthetic series that has contrast
- Dangerous to use clinically "its just made up data"
- useful for developers - access to real world data is difficult
- Concern about this generative data be mixed in with real data and polluting ground truth
- DICOM Tags to help with this
- Tie in with ontology/procedure codes - could this be used to help define terms, procedures?
- Very useful for developers as they can capture the key features that define a procedure type (statistically equivalent)
- DICOM Tags for manufacturer, model number as input?
- Question - how does one create a generative ai for medical images?
-
OpenAI "Whisper" (audio to text in 40 languages)
- OpenAI requires a license to use - $20EU/month for API access?
- https://github.com/openai/whisper
- https://platform.openai.com/docs/api-reference/audio/create
- Weights were released? Whisper.cpp (open source?) - FREE
- How well does it work with a radiology report?
- Did ok, but replaced some terms with "more commmon" terms that sound similar
- Perhaps some way to train this
- used
- Can run without a GPU, WASM version in browser
-
Microsoft Office 365 Copilot
- https://www.youtube.com/watch?v=S7xTBa93TX8
- Google has generative AI for google workplaces: https://twitter.com/benparr/status/1635684322261729282?s=12&t=sFpjko2Fp7Yl8nMqjSK8Xg
-
Clinical Trial Data Modeling
- Clinical Trial
- 1:N Site
- 1:N Subjects
- 1:N Visit (baseline, week 4, etc)
- 1:N Time Point (collection of data at a visit)
- 1:N DICOM Studies
- OMOP - should this be used for clincial trials