AI alignment in medical imaging: Unveiling hidden biases through counterfactual analysis
PLOS Digital Health · 13.08.2026 · 5 (8) · e0001516
DOI: 10.1371/journal.pdig.0001516 ↗Kurz gefasst
Fachartikel in der Zeitschrift PLOS Digital Health, online erschienen am 13.08.2026. Verfasst von Haroui Ma und 3 weiteren Autor:innen. Beteiligte Einrichtung: TUM School of Computation, Information and Technology (CIT), Technical University Munich, Munich, Germany. Die Arbeit nennt 64 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Helmholtz Foundation Model Initiative, Gauss Centre for Supercomputing.
Abstract (Originalsprache)
Machine learning (ML) systems for medical imaging have demonstrated remarkable diagnostic capabilities. However, their susceptibility to learning spurious correlations with sensitive attributes poses significant risks to fairness and safety. In this paper, we introduce a novel statistical framework to evaluate the dependency of medical imaging ML models on sensitive attributes, such as demographics. Our method leverages the concept of counterfactual invariance, measuring the extent to which a model’s predictions remain unchanged under hypothetical changes to sensitive attributes. We present a practical algorithm that combines conditional latent diffusion models with statistical hypothesis testing to identify and quantify such biases without requiring direct access to counterfactual data. On synthetic benchmarks, our framework correctly identifies 96.1% of biased models while producing false alarms on 22.0% of fair models. On two real-world chest X-ray datasets, CheXpert and MIMIC-CXR, it detects bias at average rates 96.3%, 95.7% across all diagnostic tasks when models are strongly biased, and its average false alarm rate drops to 15.3%, 14.7%, respectively, for the least biased models, consistently outperforming existing fairness baselines and demonstrating strong alignment with counterfactual fairness principles.
Autor:innen und Institutionen
- TUM School of Computation, Information and Technology (CIT), Technical University Munich, Munich, Germany.
- Department of Engineering Science, The University of Oxford, Oxford, United Kingdom.
- TUM School for Medicine and Health, Institute of History and Ethics in Medicine, Technical University of Munich, Munich, Germany.
- Helmholtz AI, Helmholtz Munich, Munich, Germany.
ORCID-Profile: Francesco Quinzan
Förderung
- Helmholtz Foundation Model Initiative
- Gauss Centre for Supercomputing
Zitieren
APA
Ma, H., Quinzan, F., Willem, T., & Bauer, S. (2026). AI alignment in medical imaging: Unveiling hidden biases through counterfactual analysis. PLOS Digital Health, 5(8), e0001516. https://doi.org/10.1371/journal.pdig.0001516
Vancouver
Ma H, Quinzan F, Willem T, Bauer S. AI alignment in medical imaging: Unveiling hidden biases through counterfactual analysis. PLOS Digital Health. 2026;5(8):e0001516. doi:10.1371/journal.pdig.0001516
BibTeX
@article{Ma2026,
author = {Ma, Haroui and Quinzan, Francesco and Willem, Theresa and Bauer, Stefan},
title = {AI alignment in medical imaging: Unveiling hidden biases through counterfactual analysis},
journal = {PLOS Digital Health},
year = {2026},
volume = {5},
number = {8},
pages = {e0001516},
doi = {10.1371/journal.pdig.0001516}
}
EUVIMED