Data, data everywhere
- data is abundant, good data is scarce
- ethical guidelines - minimum data vs all data
- imaging biobanks
- what is good quality data
- what do we need for applying ML
- risk of overfitting, performance on cross-validation
- clinical data is heterogeneous, but ML works best on clean, homogeneous data
- how to assess data quality, the concept of data readiness levels
- what can clinicians do, how to label/annotate/curate data (mechanical turk, structured reporting)
- how to validate ML, how to get approval (is this a different paper?)
Co-design of ML imaging systems
- domain expertise is essential in building ML systems
- interdisciplinary teams
- experts in the loop from design until deployment
- learning the right features
- end-to-end learning (as in neural networks) comes with the risk of wrong data associations