-
Notifications
You must be signed in to change notification settings - Fork 1.1k
[2020.12.10] Deploy WG: KickOff
- Ralf (DKFZ) - 2. Klaus Kades (DKFZ) - 3. Jonas (DKFZ) - 4. Marco (DKFZ) - 5. Daniel (Stanford) - 6. Bennett (Vanderbilt) - 7. Jayashree (MGH) - 8. Anthony (Mt. Sinai) - 9. Haris (GST) - 10. Krishna (NIH/MSKCC) - 11. Jorge (KCL) - 12. Selnur (Mayo Jacksonville) - 13. Stephen (Kitware) - 14. Risto (NVIDIA) - 14. Rahul (NVIDIA) - 15. Dana (NVIDIA) - 16. Michael (NVIDIA) - 17. David (NVIDIA)
12/10/2020 - MONAI Deploy WG kick-off.mp4
- AI1 - Publish notes and scope publicly (DB).
- AI2 - Identify what WGs are stakeholders. Map out dependencies (JC).
- AI3 - Create core WG. (sign up!). Meet 60 mins weekly or 90 mins bi-weekly.
60 mins - Once a month, starting Jan.
- Clinical implementation + interoperability (starting with DICOM, then EHR).
- Central repository to facilitate collaboration among institutions.
- By March 1st, first draft roadmap.
- By April 1st, final 2021 roadmap.
- By June 1st, API definition and first demo/prototype.
What other WGs are stakeholders/we depend on.
-
[Jorge] How far in the stack we want MONAI Deploy to be - definition vs implementation Imaging first EHR/EMR next General purpose Data pipelines from different data sources GPU accelerated computing Safe for students development algorithms without deep expertise Apache 2 OSS license Model Zoos Sharing models/AI apps between institutions at a later phase Check with MONAI Core or other WGs where things could live: AI JC
-
[Bennett] Reference prototype/implementation ASAP to shift ongoing efforts
-
[Stephen] Integration from trained models to clinical integration (more interesting than research workflows)
-
[Ralf] Need balance between definition and jumping too quick to implementation To avoid discussing to much to find the “one” implementation Some institutions already have their own implementations Clear definition of the scope of the API Need DICOM Striving for something like a connectathon, so MONAI Deploy and its API definition make it easy for us to provide interoperability of our development. Potential collaboration with the benchmarking team.
-
[Selnur] +1 clinical integration. Connectivity is key: DICOM, SR, SEG HL7/FHIR Model has to be already packaged within the inference workflow, ready to go, with an end-to-end example/pipeline. Little test. Fast team 6-8 months from model ready from research until it is deployment ready
-
[Anthony] ML Platform - not imaging specific +1 clinical integration
-
[Haris] +1 clinical integration Source Auditing Execution Simplify for scientists so they don’t need to understand DICOM and deep dive into modalities ModelZoo
-
[Stephen] RSNA compliance end-to-end test as validation check. Look at DICOM/CTN history as inspiration Was used to bring DICOM vendors into compliance: https://www.mir.wustl.edu/Portals/0/Documents/Uploads/ERL/overview.pdf A model should produce similar outcomes on clinical AI inference systems provided by multiple vendors for the model and for each vendor to be compliant with MONAI Deploy standard
-
[Rahul] Receiving DICOM is not difficult, standardized data input to the model is the tough part. Pipeline can include different models Orchestration of the models Performance at clinical inference is more difficult and important than in training Debuggability Monitor Model drift and what to do then Visualization friendly output to PACS
- What’s out there already that we like and can reuse?
- Vanderbilt U. - ModelZoo
- Vanderbilt has been harmonizing API’s and has a working draft of a consensus protocol for wrapping models as part of an NSF convergence project. We can demo / share working documents and a pilot demonstration. Formal documentation is ongoing and due by Feb. 2021 (-Bennett).
- Infrastructure-Stack | Clinical Data Science | Mount Sinai
- What do we need to create ourselves?
- Who would create it?
- Who are our users?
- WIP after conversation with Anthony here: MONAI Deploy User Personas