MIMM-X: Disentangling Spurious Correlations for Medical Image Analysis

2025-11-28

AI Summary

In medical image analysis, deep learning models are prone to spurious correlations—so-called “shortcut learning”—across multiple confounding factors, leading to poor cross-domain generalization and elevated clinical risks. To address this, we propose MIMM-X, the first framework to jointly disentangle multiple spurious associations in multimodal medical imaging (MRI + X-ray) via mutual information minimization, explicitly separating causal pathological features from confounding factors. Integrating causal representation learning with multi-source data joint modeling, MIMM-X requires no auxiliary annotations or domain labels. Evaluated on three large-scale public datasets—UK Biobank, NAKO, and CheXpert—MIMM-X significantly mitigates spurious correlations, enhances cross-center and cross-device generalization, and achieves average AUC improvements of 3.2–5.7 percentage points. Our approach establishes a novel, interpretable, and scalable paradigm for robust, causally grounded medical AI.

Technology Category

• Machine Learning: Multimodal Learning
• Computer Vision: Multi-modal Vision
• Reasoning under Uncertainty: Causality

Application Category

• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

Deep learning models can excel on medical tasks, yet often experience spurious correlations, known as shortcut learning, leading to poor generalization in new environments. Particularly in medical imaging, where multiple spurious correlations can coexist, misclassifications can have severe consequences. We propose MIMM-X, a framework that disentangles causal features from multiple spurious correlations by minimizing their mutual information. It enables predictions based on true underlying causal relationships rather than dataset-specific shortcuts. We evaluate MIMM-X on three datasets (UK Biobank, NAKO, CheXpert) across two imaging modalities (MRI and X-ray). Results demonstrate that MIMM-X effectively mitigates shortcut learning of multiple spurious correlations.

Problem

• Disentangling causal features from spurious correlations in medical images
• Mitigating shortcut learning to improve generalization across datasets
• Reducing misclassifications from multiple coexisting spurious correlations

Innovation

• Disentangles causal features from spurious correlations
• Minimizes mutual information to prevent shortcut learning
• Enables predictions based on true causal relationships