FairVision: Equitable Deep Learning for Eye Disease Screening via Fair Identity Scaling

TL;DR

The paper tackles the underexplored fairness of 3D medical imaging by auditing eye-disease screening models across race, gender, and ethnicity, using both 2D and 3D data. It introduces Fair Identity Scaling (FIS), a loss weighting framework combining individual and group equity via learned loss weights and optimal transport, to improve both accuracy and fairness. The authors provide theoretical guarantees for fair learning and demonstrate through Harvard-FairVision, a 30,000-subject dataset with 2D and 3D imaging and six demographics, that FIS outperforms SOTA fairness methods while boosting overall AUC. This work advances practical deployment of fair AI in ophthalmology by delivering a large dataset and a scalable fairness mechanism with proven benefits. The Harvard-FairVision resource and the FIS method enable more equitable eye-disease screening in diverse populations, potentially reducing health disparities.

Abstract

Equity in AI for healthcare is crucial due to its direct impact on human well-being. Despite advancements in 2D medical imaging fairness, the fairness of 3D models remains underexplored, hindered by the small sizes of 3D fairness datasets. Since 3D imaging surpasses 2D imaging in SOTA clinical care, it is critical to understand the fairness of these 3D models. To address this research gap, we conduct the first comprehensive study on the fairness of 3D medical imaging models across multiple protected attributes. Our investigation spans both 2D and 3D models and evaluates fairness across five architectures on three common eye diseases, revealing significant biases across race, gender, and ethnicity. To alleviate these biases, we propose a novel fair identity scaling (FIS) method that improves both overall performance and fairness, outperforming various SOTA fairness methods. Moreover, we release Harvard-FairVision, the first large-scale medical fairness dataset with 30,000 subjects featuring both 2D and 3D imaging data and six demographic identity attributes. Harvard-FairVision provides labels for three major eye disorders affecting about 380 million people worldwide, serving as a valuable resource for both 2D and 3D fairness learning. Our code and dataset are publicly accessible at \url{https://ophai.hms.harvard.edu/datasets/harvard-fairvision30k}.

Figures (4)

• Figure 1: Illustration of comparison between conventional supervised learning and fair supervised learning. Conventional supervised learning is prone to neglecting underrepresented groups' performance. In contrast, fair supervised learning aims to enhance underrepresented groups' performance while retaining or even improving overrepresented groups' performance.
• Figure 2: (a) 3D OCT B-scans. (b) SLO fundus image. (c) The label distribution for AMD, (d) The label distribution for DR, and (e) The label distribution for glaucoma.
• Figure 3: Schematic view of the proposed FIS, in comparison to the standard supervised learning framework. Given inputs such as SLO fundus images or OCT B-scans, the proposed FIS incorporates individual scaling and group scaling to determine adaptive loss weights. In the visualization, a large circle in the loss (or loss weight) indicates a high value for the instance loss (or loss weight).
• Figure 4: Effects of the fusion weight $c$ on AUC and Mean PSD in AMD detection (left), DR detection (middle), and glaucoma detection (right). SLO fundus images are used for this analysis.

Theorems & Definitions (16)

• Lemma 4.1
• proof
• Theorem 4.2
• proof
• Theorem 4.3
• proof
• Theorem 4.4
• proof
• Theorem 4.5
• proof