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

Shaojin Bai; Yuting Su; Weizhi Nie

cs.CVMar 26, 2026

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

Authors: Shaojin Bai, Yuting Su, Weizhi Nie

Organizations: School of Electrical and Information Engineering, Tianjin University, Tianjin, China

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.

Figures & tables

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.

Method	Camelyon17			CXR
Method	Acc ( $\uparrow$ )	Wg-Acc ( $\uparrow$ )	ECE ( $\downarrow$ )	AUC ( $\uparrow$ )	EOD ( $\downarrow$ )	DPD ( $\downarrow$ )
SWAD [ 23 ]	$92.45\pm 0.4$	$67.82\pm 1.2$	$10.15\pm 0.5$	$82.64\pm 0.3$	$8.42\pm 0.4$	$9.15\pm 0.4$
LISA [ 24 ]	$92.18\pm 0.5$	$68.94\pm 1.1$	$9.84\pm 0.6$	$82.91\pm 0.3$	$8.15\pm 0.4$	$8.86\pm 0.5$
DFR [ 25 ]	$91.92\pm 0.6$	$69.85\pm 1.0$	$9.62\pm 0.5$	$83.12\pm 0.2$	$7.94\pm 0.3$	$8.53\pm 0.4$
CLIP (Zero-Shot) [ 26 ]	$85.23\pm 0.5$	$52.16\pm 1.8$	$18.52\pm 1.2$	$79.54\pm 0.0$	$14.81\pm 0.0$	$16.25\pm 0.0$
MedCLIP [ 27 ]	$84.58\pm 0.6$	$50.82\pm 1.5$	$19.24\pm 1.1$	$81.23\pm 0.2$	$11.56\pm 0.3$	$12.42\pm 0.4$

Table 1: Main Results: Comparison with State-of-the-Art Methods. All scores presented in percentage. The best results are highlighted in bold.

Method	Components		Camelyon17			CXR
Method	IV	Cond.	Acc ( $\uparrow$ )	Wg-Acc ( $\uparrow$ )	ECE ( $\downarrow$ )	AUC ( $\uparrow$ )	EOD ( $\downarrow$ )	DPD ( $\downarrow$ )
ERM (Baseline)	$-$	$-$	$88.50\pm 0.6$	$55.20\pm 1.5$	$15.42\pm 0.8$	$80.12\pm 0.4$	$12.45\pm 0.5$	$13.10\pm 0.5$
DeepGMM (w/o CIV)	✓	$-$	$91.25\pm 0.5$	$63.54\pm 1.2$	$11.82\pm 0.6$	$81.43\pm 0.3$	$10.12\pm 0.4$	$10.85\pm 0.5$
CIV-DG (Random $Z$ )	✓	Rand	$92.10\pm 0.4$	$66.85\pm 1.1$	$10.05\pm 0.5$	$82.50\pm 0.3$	$8.95\pm 0.4$	$9.20\pm 0.4$
CIV-DG	✓	Causal	$\mathbf{93.24\pm 0.3}$	$\mathbf{70.46\pm 0.9}$	$\mathbf{8.73\pm 0.3}$	$\mathbf{84.15\pm 0.2}$	$\mathbf{6.81\pm 0.3}$	$\mathbf{6.78\pm 0.3}$

Table 2: Ablation study on CIV-DG components. All scores presented in percentage. The best results are highlighted in bold.

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.

Explore similar work

Causal Transfer in Medical Image Analysis
Mar 25, 2026Mohammed M. Abdelsamea, Daniel Tweneboah Anyimadu, Tasneem Selim +4Medical Image Analysis Cross-Domain Transfer Learning
Department of Computer Science, University of Exeter, Exeter, United Kingdom · Department of Mathematics and Computer Science, Faculty of Science, Alexandria University, Alexandria, Egypt · Hawkes Institute, Department of Computer Science, University College London, London, United Kingdom
CIPHER: Causal Intervention Pathways for Healthcare Equity and Robustness
Jul 1, 2026Xinyu Jia, Weidong Guo, Wangyuan Zhao +3Algorithmic Fairness Medical Image Synthesis
College of Biomedical Engineering, Fudan University, Shanghai, China
CoDG-Net: Structure-Guided Style Diffusion and Collaborative Learning to Mitigate Catastrophic Forgetting in Medical Image Domain Generalization
Oct 4, 2026Yucheng Song, Jincan Wang, Haokang Ding +3Data Augmentation Cross-Domain Generalization
School of Computer Science and Engineering, Central South University, Changsha, China · School of Software Engineering, Xi’an Jiaotong University, Xi’an, Shaanxi, China · College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China

Mar 25, 2026J/K move · Enter open · S save

Causal Transfer in Medical Image Analysis

Mohammed M. Abdelsamea, Daniel Tweneboah Anyimadu, Tasneem Selim, Saif Alzubi+3

Department of Computer Science, University of Exeter, Exeter, United Kingdom · Department of Mathematics and Computer Science, Faculty of Science, Alexandria University, Alexandria, Egypt · Hawkes Institute, Department of Computer Science, University College London, London, United Kingdom

Medical Image Analysis Cross-Domain Transfer Learning Transfer Learning Causal Inference

Open paper AskPDF arXiv

Medical imaging models frequently fail when deployed across hospitals, scanners, populations, or imaging protocols due to domain shift, limiting their clinical reliability. While transfer learning and domain adaptation address such shifts statistically, they often rely on spurious correlations that break under changing conditions. On the other hand, causal inference provides a principled way to identify invariant mechanisms that remain stable across environments. This survey introduces and systematises Causal Transfer Learning (CTL) for medical image analysis. This paradigm integrates causal reasoning with cross-domain representation learning to enable robust and generalisable clinical AI. We frame domain shift as a causal problem and analyse how structural causal models, invariant risk minimisation, and counterfactual reasoning can be embedded within transfer learning pipelines. We studied spanning classification, segmentation, reconstruction, anomaly detection, and multimodal imaging, and organised them by task, shift type, and causal assumption. A unified taxonomy is proposed that connects causal frameworks and transfer mechanisms. We further summarise datasets, benchmarks, and empirical gains, highlighting when and why causal transfer outperforms correlation-based domain adaptation. Finally, we discuss how CTL supports fairness, robustness, and trustworthy deployment in multi-institutional and federated settings, and outline open challenges and research directions for clinically reliable medical imaging AI.

Abstract

Explore similar work

Causal Transfer in Medical Image Analysis

CIPHER: Causal Intervention Pathways for Healthcare Equity and Robustness

CoDG-Net: Structure-Guided Style Diffusion and Collaborative Learning to Mitigate Catastrophic Forgetting in Medical Image Domain Generalization