The whole truth about using deep learning in medical imaging
summary
The talk presents replacing legacy patch-based segmenters with a U-Net to cut brain MRI inference time from about 10 minutes to 1 second while maintaining similar segmentation quality. It introduces a dataset-wide binary adjacency matrix of which anatomical regions touch and adds an adjacency-penalty term to the loss so networks learn to avoid anatomically impossible contacts; evaluation on mika, bsr and anatomy3 shows fewer impossible adjacencies and improved robustness for clinical deployment.
AI-written summary · about
chapters (23)
- 0:00Introduction
- 3:48Understanding the human body
- 6:25Multiple ways to see
- 8:04Why so many possibilities?
- 9:24Multiple ways to "see"
- 11:13Some challenges in imaging
- 12:36Inventing tomorrow's imaging tools
- 16:09Deep learning in medical imaging
- 18:06Supervised learning
- 19:14Medical image segmentation
- 20:40CNN architectures for segmentation
- 21:52Challenges in medical segmentation
- 24:32To begin with, you need a problem
- 27:50Think like a doctor
- 28:28Modeling human knowledge
- 31:12Binary adjacency extraction
- 32:48Integrating the adjacency constraint
- 34:13Results after integrating the constraint
- 35:43And once the work is done
- 36:15Some French startups
- 39:10Cerebral Vascular Accident
- 39:30Treatments for stroke
- 40:04Automatic visualization for stroke
related talks
43mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
49mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
33mUnsupervised and Semi-Supervised Deep Learning for Medical Imaging
42mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
34mWhy AI in the Medical Field Both Sucks and Rocks...
50mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
27mU-Net is all you need: applying the same AI approach for many different business domains
25mDeep Learning in Biomedicine: Challenges and Opportunities