FairCLIP: Harnessing Fairness in Vision-Language Learning

TL;DR

This work tackles fairness in medical vision-language models by introducing Harvard-FairVLMed, a large, de-identified dataset pairing fundus images with clinical notes and rich demographic attributes for fairness analysis. It benchmarks CLIP and BLIP2 across natural and medical pretraining, revealing persistent biases across race, gender, ethnicity, and language. To address these biases, the authors propose FairCLIP, a Sinkhorn-distance-based fairness objective that aligns the overall sample distribution with subgroup distributions during training, yielding improved ES-AUC and reduced disparity metrics. The dataset and code are released to catalyze development of ethically aware and clinically effective VL models, with extensive ablations supporting the robustness of the FairCLIP approach.

Abstract

Fairness is a critical concern in deep learning, especially in healthcare, where these models influence diagnoses and treatment decisions. Although fairness has been investigated in the vision-only domain, the fairness of medical vision-language (VL) models remains unexplored due to the scarcity of medical VL datasets for studying fairness. To bridge this research gap, we introduce the first fair vision-language medical dataset Harvard-FairVLMed that provides detailed demographic attributes, ground-truth labels, and clinical notes to facilitate an in-depth examination of fairness within VL foundation models. Using Harvard-FairVLMed, we conduct a comprehensive fairness analysis of two widely-used VL models (CLIP and BLIP2), pre-trained on both natural and medical domains, across four different protected attributes. Our results highlight significant biases in all VL models, with Asian, Male, Non-Hispanic, and Spanish being the preferred subgroups across the protected attributes of race, gender, ethnicity, and language, respectively. In order to alleviate these biases, we propose FairCLIP, an optimal-transport-based approach that achieves a favorable trade-off between performance and fairness by reducing the Sinkhorn distance between the overall sample distribution and the distributions corresponding to each demographic group. As the first VL dataset of its kind, Harvard-FairVLMed holds the potential to catalyze advancements in the development of machine learning models that are both ethically aware and clinically effective. Our dataset and code are available at https://ophai.hms.harvard.edu/datasets/harvard-fairvlmed10k.

Figures (5)

• Figure 1: Examples of non-glaucomatous and glaucomatous samples with the corresponding SLO fundus image and clinical note.
• Figure 2: Schematic view of the proposed FairCLIP. Clinical notes containing PHI (e.g., names and gender) undergo de-identification and summarization to fit text encoder limitations, such as CLIP's 77-token maximum length. FairCLIP equalizes the overall sample distribution with the distributions corresponding to each demographic group, thereby achieving a favorable trade-off between performance and fairness.
• Figure 3: Extensive analyses on Harvard-FairVLMed, including (a) Effects of various LLM summarizations on the performance-fairness trade-off of BLIP2, (b) Effects of using pre-trained vision encoders from natural (CLIP) and medical (PMC-CLIP) domains on the performance-fairness trade-off of BLIP2, and (c) Performance comparison among CLIP, CLIP with adversarial loss (CLIP w/ Adv), and FairCLIP.
• Figure 4: (a) Distribution of words in the clinical notes, (b) Prevalence of subjects across race, (c) Prevalence of subjects across gender.
• Figure 5: (a) Ablation study of using various $|\mathcal{B}_{a}|$ in FairCLIP, (b) Ablation study on the effects of $\epsilon$ on model performance, (c) Fairness results based on marital status, (d) Comparison of FairCLIP against other fairness algorithms.