Project @ UBRA AI Toolbox Hackathon
- Create a GitHub account (if you do not have one yet)
- Fork this repository and give all group members access to the fork; I suggest using this as the primary code synchronization method
- Create a Weights & Biases account (if you do not have one yet - one per group is sufficient)
- Take notice of the CheXpert dataset research use agreement and create an individual account here
- Get onto the VM and verify you can run cxp_pneu.py
- Go through the baseline code; make sure you understand it
- Implement the Serna et al. approach as a potential shortcut learning mitigation method
- Go through the list of open issues; address what you find interesting or explore your own ideas freely
General shortcut learning:
- Shortcut learning in deep neural networks
- The risk of shortcutting in deep learning algorithms for medical imaging research
Papers that cover pneumothorax/chest drain shortcut learning:
- Hidden stratification causes clinically meaningful failures in machine learning for medical imaging
- DETECTING SHORTCUTS IN MEDICAL IMAGES - A CASE STUDY IN CHEST X-RAYS
- Slicing Through Bias: Explaining Performance Gaps in Medical Image Analysis using Slice Discovery Methods
Drawbacks of DANN / CDANN (baseline method for 'domain invariance' which can also be used to address shortcut learning):
- Fundamental Limits and Tradeoffs in Invariant Representation Learning
- 10 Years of Fair Representations: Challenges and Opportunities
- Are demographically invariant models and representations in medical imaging fair?
- The limits of fair medical imaging AI in real-world generalization
- MEDFAIR: Benchmarking Fairness for Medical Imaging
Alternative approach pursued here: Sensitive loss: Improving accuracy and fairness of face representations with discrimination-aware deep learning