[Paper Review] FairMedFM: Fairness Benchmarking for Medical Imaging Foundation Models

Ruinan Jin, Zikang Xu+5 more|arXiv (Cornell University)|Jul 1, 2024

Health Systems, Economic Evaluations, Quality of Life4 citations

TL;DR

FairMedFM introduces a comprehensive fairness benchmark for medical imaging foundation models (FMs), integrating 17 diverse medical datasets across modalities and dimensions, and evaluating 20 FMs using multiple downstream tasks (classification and segmentation) and fairness-aware techniques. Key findings reveal persistent bias across models, significant utility-fairness trade-offs, and limited effectiveness of existing mitigation strategies, with no FM consistently outperforming others in fairness or utility.

ABSTRACT

The advent of foundation models (FMs) in healthcare offers unprecedented opportunities to enhance medical diagnostics through automated classification and segmentation tasks. However, these models also raise significant concerns about their fairness, especially when applied to diverse and underrepresented populations in healthcare applications. Currently, there is a lack of comprehensive benchmarks, standardized pipelines, and easily adaptable libraries to evaluate and understand the fairness performance of FMs in medical imaging, leading to considerable challenges in formulating and implementing solutions that ensure equitable outcomes across diverse patient populations. To fill this gap, we introduce FairMedFM, a fairness benchmark for FM research in medical imaging.FairMedFM integrates with 17 popular medical imaging datasets, encompassing different modalities, dimensionalities, and sensitive attributes. It explores 20 widely used FMs, with various usages such as zero-shot learning, linear probing, parameter-efficient fine-tuning, and prompting in various downstream tasks -- classification and segmentation. Our exhaustive analysis evaluates the fairness performance over different evaluation metrics from multiple perspectives, revealing the existence of bias, varied utility-fairness trade-offs on different FMs, consistent disparities on the same datasets regardless FMs, and limited effectiveness of existing unfairness mitigation methods. Checkout FairMedFM's project page and open-sourced codebase, which supports extendible functionalities and applications as well as inclusive for studies on FMs in medical imaging over the long term.

Motivation & Objective

• Address the lack of standardized, comprehensive benchmarks for evaluating fairness in medical imaging foundation models (FMs).
• Systematically assess fairness across diverse FMs, data modalities, sensitive attributes (e.g., sex, age, race), and downstream tasks (classification and segmentation).
• Investigate the trade-offs between model utility and fairness across different fine-tuning strategies and FM types.
• Provide an extensible, open-source codebase to support long-term research on fairness in medical FMs.
• Evaluate the effectiveness of existing fairness mitigation techniques in real-world medical imaging scenarios.

Proposed method

• Integrates 17 medical imaging datasets spanning 2D, 2.5D, and 3D modalities (X-ray, CT, MRI, ultrasound, OCT, fundus, dermatoscopy).
• Supports 20 widely used FMs, including general-purpose (e.g., CLIP, ViT) and medical-specific models (e.g., Med-PaLM, SAM-Med3D).
• Enables multiple usage paradigms: zero-shot, linear probing, parameter-efficient fine-tuning (PEFT), prompt-based adaptation, and full fine-tuning.
• Employs a unified codebase built on PyTorch with modular components: Dataloader, Model, Usage Wrapper, Trainer, and Evaluation modules.
• Applies diverse fairness metrics including EqOdds, AUCΔ, ECEΔ, DSCΔ, and predictive-alignment fairness, with representation fairness analyzed via t-SNE.
• Incorporates state-of-the-art fairness mitigation techniques such as group rebalancing, adversarial training, and fairness constraints.

Figure 1: Overview of the FairMedFM framework, a standardized pipeline to investigate fairness on diverse datasets (2D, 2.5D, and 3D), comprehensive functionalities (various FMs, tasks, usages, and debias algorithms), thorough evaluation metrics. The details are explained in Sec. 3 .

Experimental results

Research questions

• RQ1 How do different foundation models perform in terms of fairness across diverse medical imaging datasets and sensitive attributes?
• RQ2 What is the nature and extent of the utility-fairness trade-off across various FM types, usage strategies, and downstream tasks?
• RQ3 To what extent do existing fairness mitigation techniques reduce bias in medical imaging FMs?
• RQ4 Are fairness disparities consistent across different FMs when applied to the same dataset?
• RQ5 How does the latent representation of FMs relate to fairness outcomes, and can representation space manipulation mitigate bias?

Key findings

• No foundation model consistently outperforms others in fairness or utility, as confirmed by statistical tests (CD diagrams) showing no significant performance differences across models.
• Significant fairness disparities exist across sensitive attributes (e.g., sex, age, race) regardless of the FM used, indicating persistent bias in model behavior.
• The utility-fairness trade-off is pronounced and varies across FM types and usage strategies, with no single approach achieving optimal balance.
• Existing fairness mitigation methods—such as adversarial training and group rebalancing—show limited effectiveness in reducing bias across datasets and tasks.
• Representation space analysis reveals that data-level group separability (e.g., in HAM10000) correlates with higher fairness disparities, suggesting data distribution as a key source of bias.
• FairMedFM’s open-source codebase enables extensible benchmarking and supports diverse use cases, including segmentation with prompt-based models like SAM-Med2D and FastSAM-3D.

Figure 2: Bias in classification tasks. $\text{AUC}_{\Delta}$ is the fairness evaluation metric. “ $\dagger$ ” denotes vision-language models, and “ $*$ ” denotes pure vision models, where CLIP-ZS and CLIP-Adapt are not applicable.