Fair Foundation Models for Medical Image Analysis: Challenges and Perspectives

TL;DR

This paper tackles fairness in medical imaging foundation models, highlighting the risk that such models could deepen health disparities if bias is not addressed. It advocates a comprehensive, pipeline-wide bias mitigation framework that spans data documentation, data curation, training, evaluation, deployment, and policymaker engagement, arguing that isolated model-level fixes are insufficient. Key contributions include a global assessment of dataset representation, a conceptual framework for integrated bias mitigation, and discussion of policy interventions and equitable resource allocation to support fair adoption. The work emphasizes the potential of generative data and knowledge-aggregation approaches to broaden population representation, while underscoring the need for governance, transparency, and collaborative action to ensure the benefits of fair FMs reach underserved regions and communities.

Abstract

Ensuring equitable Artificial Intelligence (AI) in healthcare demands systems that make unbiased decisions across all demographic groups, bridging technical innovation with ethical principles. Foundation Models (FMs), trained on vast datasets through self-supervised learning, enable efficient adaptation across medical imaging tasks while reducing dependency on labeled data. These models demonstrate potential for enhancing fairness, though significant challenges remain in achieving consistent performance across demographic groups. Our review indicates that effective bias mitigation in FMs requires systematic interventions throughout all stages of development. While previous approaches focused primarily on model-level bias mitigation, our analysis reveals that fairness in FMs requires integrated interventions throughout the development pipeline, from data documentation to deployment protocols. This comprehensive framework advances current knowledge by demonstrating how systematic bias mitigation, combined with policy engagement, can effectively address both technical and institutional barriers to equitable AI in healthcare. The development of equitable FMs represents a critical step toward democratizing advanced healthcare technologies, particularly for underserved populations and regions with limited medical infrastructure and computational resources.

Figures (3)

• Figure 1: Conceptual framework. This figure delineates sequential phases in FMs development, illustrating principal challenges for achieving fairness. Three focal domains are emphasized: data documentation, curation dataset ensuring diversity and quality to detect and mitigate bias; environmental impact, spanning training, deployment, and resource efficiency; and policymakers, establish governance, standards, and resource distribution to ensure ethical, equitable FMs access.
• Figure 2: Global distribution of medical imaging data: This geographic visualization depicts the volume of medical imaging datasets by country, excluding multi-country datasets listed in Table \ref{['tab:medical-datasets']}. The figure highlights pronounced disparities in data representation, underscoring its critical role in the development of equitable FMs. Notably, a limited number of countries account for the majority of available medical imaging data.
• Figure 3: Overview recommendation: The figure summarizes recommendations for achieving fair foundation models, indicating for each stage its medical infrastructure, importance for fairness (essential), high cost, and computation. The recommendations focus on where governments and industry should prioritize technical interventions to advance fairness. Medical infrastructure is critical for both downstream applications and data creation, but their cost profiles differ: data creation depends primarily on expensive imaging equipment and large-scale data storage, whereas downstream tasks require substantial clinical expertise for data annotation. With respect to fairness mitigation, the most critical and challenging stage is pre-training, which demands extensive computational infrastructure and incurs high costs. The figure also highlights evaluation and data curation as comparatively low-cost yet important strategies for improving fairness. Generative models offer a promising avenue to balance data creation and alleviate infrastructure constraints by better representing the underlying population. At the same time, real-data collection remains more critical for fairness, because equitable access to medical infrastructure across all population groups is essential, whereas generative models can only complement these efforts by improving population representation in the training data. Finally, agglomerative models and world models are identified as future directions with potential to enhance fairness, although they are tightly coupled to pre-training and therefore are not quantitatively assessed in the figure.