Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CIV-DG: Conditional Instrumental Variables for Domain Generalization in Medical Imaging

Shaojin Bai, Yuting Su, Weizhi Nie

Abstract

Cross-site generalizability in medical AI is fundamentally compromised by selection bias, a structural mechanism where patient demographics (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. To surmount this identifiability barrier, we propose CIV-DG, a causal framework that leverages Conditional Instrumental Variables to disentangle pathological semantics from scanner-induced artifacts. By relaxing the strict random assignment assumption of standard IV methods, CIV-DG accommodates complex clinical scenarios where hospital selection is endogenously driven by patient demographics. We instantiate this theory via a Deep Generalized Method of Moments (DeepGMM) architecture, employing a conditional critic to minimize moment violations and enforce instrument-error orthogonality within demographic strata. Extensive experiments on the Camelyon17 benchmark and large-scale Chest X-Ray datasets demonstrate that CIV-DG significantly outperforms leading baselines, validating the efficacy of conditional causal mechanisms in resolving structural confounding for robust medical AI.

CIV-DG: Conditional Instrumental Variables for Domain Generalization in Medical Imaging

Abstract

Cross-site generalizability in medical AI is fundamentally compromised by selection bias, a structural mechanism where patient demographics (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. To surmount this identifiability barrier, we propose CIV-DG, a causal framework that leverages Conditional Instrumental Variables to disentangle pathological semantics from scanner-induced artifacts. By relaxing the strict random assignment assumption of standard IV methods, CIV-DG accommodates complex clinical scenarios where hospital selection is endogenously driven by patient demographics. We instantiate this theory via a Deep Generalized Method of Moments (DeepGMM) architecture, employing a conditional critic to minimize moment violations and enforce instrument-error orthogonality within demographic strata. Extensive experiments on the Camelyon17 benchmark and large-scale Chest X-Ray datasets demonstrate that CIV-DG significantly outperforms leading baselines, validating the efficacy of conditional causal mechanisms in resolving structural confounding for robust medical AI.

Paper Structure

This paper contains 11 sections, 4 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Structural Causal Model (SCM)
  4. DeepGMM for CIV-DG
  5. Experiment & Analysis
  6. Datasets
  7. Experimental Setup
  8. Comparisons with the state-of-the-art methods
  9. Ablation studies
  10. Visualization and Discussion
  11. Conclusion

Figures (2)

  • Figure 1: Structural Causal Models (SCM) and the CIV-DG framework.(A) The generative graph illustrating how hospital assignment $Z$ induces site-specific artifacts $A$ in images $X$. (B) Our inference graph. Dashed arrows from $D$ denote methodological conditioning rather than structural causal edges. (C)The proposed CIV-DG framework. It comprises the DeepGMM Critic (top) for enforcing conditional moment restrictions via stratum-wise estimation; the Linear Representation path (bottom) using a frozen encoder and Causal Adapter for prediction; and the Optimization module (right) depicting the alternating minimax updates between the critic and primal objectives.
  • Figure 2: Qualitative analysis of CIV-DG. (A-B) t-SNE visualization showing improved feature separation over Baseline. (C) Hyperparameter sensitivity analysis of $\lambda$. (D) AUC comparison across age subgroups on CXR.