Abstract
Deep learning has become a popular tool for computer-aided diagnosis using medical images, sometimes matching or exceeding the performance of clinicians. However, these models can also reflect and amplify human bias, potentially resulting in inaccurate or missed diagnoses for certain populations. Despite this concern, the problem of improving model fairness in medical image classification has yet to be fully studied. To address this issue, we propose an algorithm that leverages marginal pairwise equal opportunity to reduce bias. Our evaluations across four tasks using four independent large-scale cohorts demonstrate that our proposed algorithm not only improves fairness in individual and intersectional subgroups but also maintains overall performance. Specifically, the relative change in pairwise fairness difference (PFD) was reduced by over 35%, while the relative change in AUC value was typically within 1%. By reducing model bias, our approach can potentially alleviate concerns about the reliability of image-based computer-aided diagnosis.
Introduction
Deep learning models for computer-aided diagnosis (CAD) have demonstrated remarkable accuracy, often rivaling human experts. However, a significant challenge remains: these models frequently inherit and amplify human biases present in training data. A model may achieve high overall accuracy while performing poorly for specific demographic groups based on race, sex, or age.
The primary obstacle in addressing this is the "fairness-accuracy trade-off." Traditional methods for improving fairness often introduce constraints that degrade a model's overall diagnostic performance. In clinical settings, any reduction in overall accuracy is difficult to justify. This research aims to achieve both group fairness (equitable performance across demographics) and intersectional fairness (e.g., performance for "young Black females") without compromising clinical utility.
Methods
The proposed approach shifts from standard Binary Cross-Entropy (BCE) loss to a training strategy based on Marginal Ranking Loss.
The methodology centers on a Pairwise Fairness Metric. Rather than simple error rates, this metric measures the probability that the model ranks a randomly selected positive sample (diseased) higher than a randomly selected negative sample (healthy) within a specific subgroup. During training, the algorithm calculates the Pairwise Fairness for every subgroup in a batch, identifies the group with the lowest fairness score, and specifically optimizes the marginal ranking loss for that group. By focusing learning efforts on the most underserved subgroups, the model improves equity without dragging down the performance of majority groups.
Datasets
The study validated this approach using four massive, independent, and publicly available datasets:
- MIDRC: Chest X-rays for COVID-19 detection (77,887 images).
- MIMIC-CXR: Chest X-rays for thorax abnormality detection (212,567 images).
- OHTS: Optic disc images for primary open-angle glaucoma (POAG) detection (37,399 images).
- AREDS: Color fundus photographs for late age-related macular degeneration (AMD) detection (66,060 images).
Key Results
The proposed method demonstrated that fairness and accuracy are not mutually exclusive. By framing fairness as a ranking problem, the researchers achieved significant improvements in equity while maintaining clinical performance.
- Significant Fairness Gains: The Pairwise Fairness Difference (PFD)—the gap between the best and worst-served groups—was reduced by over 35% compared to baseline models.
- Stable Diagnostic Performance: Despite these gains, the Area Under the Curve (AUC) remained remarkably stable, with changes typically within 1% of the baseline.
- Generalizability: The method proved effective across various neural network architectures, including DenseNet-121, DenseNet-201, and ResNet-152.
- Intersectional Improvements: The algorithm specifically improved performance in overlapping demographic categories (e.g., young females in glaucoma detection), where bias is often most pervasive.
Discussion and Future Work
While the results are promising, several areas for future research remain. The current study focuses on binary classification; extending this to multi-class diagnosis is a necessary next step. Additionally, while the model ranks patients effectively, further work is needed on probability calibration to ensure that model scores translate accurately to clinical risk across all groups. Finally, while algorithmic solutions can mitigate bias, they cannot entirely compensate for a fundamental lack of diverse data in medical research.
Summary updated