Agenda Items:
-
RSNA
- Mohannad is a rockstar!
- HyperFine is cool
- Runs on 110v
- .64 T
- Super Small
- RSNA AI/ML Stuff
- More info here: https://www.rsna.org/education/ai-resources-and-training/ai-imaging-in-practice
- Mohannad was Terry Sippel's side kick
- Premise was simple - "Imaging AI in parctice demo"
- Under the hood it was an interoperability example
- Impressions:
- Steve Borg - interoperability was highlighted. Went from philips to lunit to ?
- Mohannad -
- Startup "Within health"
- Neat followup solution
- What stood out - how they figured out the end to end problem
- communicating with patient, nagging them
- Create roadmap of actions "how to get people to do followup"
- Woojin Kim is advisor
- Startup "Within health"
- Attendance? - 40% of normal - 20,000
- Vendors said booth traffic was not effected much, interactions were good
- Fun game: https://bsbingo.techiemaestro.com/
- https://twitter.com/shadowdoc/status/1466440302626058245?s=21
-
How to embed AI results in reporting (David Kwan)
- Problem: AI in the real world doesn't work (designed around a specific cohort)
- Need metadata for AI results in standards
- Localation of scan?
- Vendor
- Product
- Version
- Mohannad - yes have thought about it somewhat, but nothing in the works yet?
- Many vendors were so new to AI/ML that they weren't familiar with the standards
- Demo does not force any vendor to do any particular thing
- Federated learning was featuree in the demo (nvidia/microsoft)
- If you are interested in joing the AI demo for RSNA 2022, talk to Mohannad Hussain
- Difficult to retrain a model and have any guarantees on performance
- e.g. tune on local cohort - cannot guarantee it has same performance if expanded to include other images
- Q: Is this question for medical devices (FDA) or from medical report perspective?
- David Kwan: Both
- Medical Devices - Point of Care (POC) US - used in ER or in surgery.
- problem with POC US, does not fit usualy diagnostic imaging workflow
- PACS is order based workflow
- IHE has a profile for encounter based workflow
- In federated learning - need to know acquisition device details (version, vendor, product, etc)
- Reporting
- if physicial takes an AI result - the metadata must be present (vendor, product, version)
- Needed in case of possible mal practice suite
- Most radiologists are having a hard time retreiving the metadata and inserting it into the report
- Radiology report is today still just text (HL7 OBX segment)
- if physicial takes an AI result - the metadata must be present (vendor, product, version)
- Medical Devices - Point of Care (POC) US - used in ER or in surgery.
- Related issue - acquisition device Q/C *phantom scanning)
- David Kwan: Both
- Q: Should reports be stored as JSON in a standard form with metadata not rendered to the rad?
- DICOM has a JSON mapping
- DICOM also has structured report IOD
- Plain text / OBX reporting is the norm in radiology, but really needs an overhaul
- Liability - who is responsible if a radiologist accepts an AI result and its wrong?
- Lung nodules - very time consuming to count/measure/report - AI can help a lot
-
Controversial subject - how important are standards in the era of web and cloud computing?
-
JPEG-XL DICOM WorkItem Approved
- NOTE - it is NOT added to the standard yet, just approved to propose it to the standard
- Main use case for JPEG-XL is to replace JPEG baseline for lossy color images
- Better compression, better color accuracy, faster
-
Q: How often is lossy compression used for diagnosis?
- BIG CAN OF WORMS
- It can be, but not allowed in some places
- used when bandwidth is limited (e.g. teleradiology, disadvantaged countries - especially less litigious)
- There are established ways to use lossy for interpretation
- Many radiologists refuse to read lossy images