Minutes Data Working Group 13 May 2020

Agenda

• Review use case scenarios, data types and flows
• Review structures for bundling data for submission

Attendees

• Invited: Brad Genereaux (Nvidia), Michael Götz (DKFZ), Carole Sudre (KCL), Stephen Aylward (Kitware), Ben Murray (KCL), Wenqi Li (Nvidia), Mona Flores (NVIDIA) Jorge Cardoso (KCL), Prerna Dogra (NVIDIA)

Notes

• Review of https://docs.google.com/spreadsheets/d/1T4W9pyaO4LqyomzQ2opdnPkFB_zuKpfhw0LQgl_4Rik/edit#gid=0
  • Scenario 1 - could sequences of MRI be used?
    • What about different data across different hospitals, or different protocols?
  • Scenario 9 - how do we represent active learning on top of existing models?
    • Correction feed on top of existing data file?
    • What happens if we don’t have the initial data that trained the model?
    • Do we need to indicate whether we have raw data or transformed data?
    • Do we need to indicate “what’s already in the model weights” and which hasn’t?
    • How do you represent the “long set of steps” for a model that has undergone active learning for quite some time?
  • What about outputs?
    • How do we tackle AI explainability?
      • Do we need to indicate which data had the most significant impact on the model?
      • Do we produce or include heat maps?
    • How would an audit trail be represented?
      • What’s the balance on what should be saved and what shouldn’t?
      • DAX (data on XNAT) might be a way

Action items

• Carole, Everyone: look at the scenarios and add/edit
• Brad, Everyone: look at the document and create more synthetic data sets
• Stephen will look into group chat (hosted by DKFZ)
• Brad: Doodle poll next week