AI RESEARCH
CIV-DG: Conditional Instrumental Variables for Domain Generalization in Medical Imaging
arXiv CS.CV
•
ArXi:2603.25202v1 Announce Type: new Cross-site generalizability in medical AI is fundamentally compromised by selection bias, a structural mechanism where patient graphics (e.g., age, severity) non-randomly dictate hospital assignment. Conventional Domain Generalization (DG) paradigms, which predominantly target image-level distribution shifts, fail to address the resulting spurious correlations between site-specific variations and diagnostic labels.