Why AI in the Medical Field Both Sucks and Rocks...
summary
Case study on segmentation of 4D heart MRI uses a U-net style model trained in tensor flow Cara's Google cloud platform on Philips scans where one patient yields ~400 images but only ~20 labeled frames. Three networks target different structures and produce Dice ~95% for left ventricle and ~58% for the right; inter-rater variability on ejection fraction is ~5%. A PyQt5 prototype and a web app demo exist, with cloud storage used and production, validation and deployment still unresolved.
AI-written summary · about
chapters (12)
- 0:00AI in medicine intro
- 3:37Cardiac MRI analysis case
- 9:18Data retrieval challenges
- 12:58Models and cloud in healthcare
- 14:25Validating AI performance
- 17:36Model to hospital deployment
- 20:06Bureaucracy barriers
- 22:55AI in medical workflows
- 23:18Doctor acceptance concerns
- 26:06Future of AI in medicine
- 27:51Call to action
- 28:44Q&A on tools and storage
related talks
49mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
18mCancer Screening using Deep Learning
43mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
50mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
42mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
37mAI in medicine: where do we stand?
27mU-Net is all you need: applying the same AI approach for many different business domains
47mThe whole truth about using deep learning in medical imaging