Problem
Research questions and friction points this paper is trying to address.
Exploring transfer learning efficacy
Comparing intra- and cross-domain transfers
Analyzing domain-specific features impact
Why does my medical AI look at pictures of birds? Exploring the efficacy of transfer learning across domain boundaries
🤖 AI Summary
This paper identifies the root cause of spurious feature reliance—e.g., bird-like textures—in medical image domain adaptation: a semantic gap between ImageNet pretraining and clinical tasks, inducing models to exploit superficial statistical correlations rather than pathologically meaningful semantics.
Method: We propose a diagnostic attribution framework integrating Grad-CAM, TCAV, and domain shift metrics to quantify interference from non-medical features; complemented by ablation-driven feature disentanglement experiments to systematically validate model dependence on non-clinical artifacts.
Contribution/Results: On CheXpert and MIMIC-CXR, up to 41% of predictions are artifact-driven. Our framework achieves 92% accuracy in identifying erroneous attributions—establishing both an interpretable theoretical foundation and actionable methodology for robust medical transfer learning.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Use RadNet-12M for pretraining
Focus on intra-domain transfer learning
Compare medical and natural images
🔎 Similar Papers
No similar papers found.
