Beyond Accuracy: Evaluating Fairness in Deep Learning Models for Chest X-ray Classification

Abstract

Deep learning (DL) models have achieved remarkable results in chest X-ray classification. However, concerns remain regarding their fairness across demographic subgroups and variations in image acquisition. Evaluations that overlook these attributes may favor models that perform well overall but exhibit biased behavior, potentially harming underrepresented subgroups. In this study, we conduct a systematic fairness evaluation of DL models trained to detect 14 clinical findings in chest radiographs. Using three public datasets and EfficientNet-Based architectures, we assess both model performance and equity across sex and age groups, as well as different radiographic views, using specialized fairness metrics. To support this analysis, we developed an evaluation framework for multi-label fairness auditing, enabling direct comparison across different models, classes, and attributes. Our results reveal notable biases even in top-performing classifiers, particularly related to patient age and positioning during image acquisition, with class-dependent bias patterns. This demonstrates that performance metrics alone do not fully characterize model behavior or the risks associated with deployment in real-world scenarios. These findings highlight the need to incorporate fairness considerations when developing and selecting models for clinical use.

Ähnliche Arbeiten