Improving diagnosis and interpretability in chest X-rays through federated learning and adversarial training

Abstract

Chest X-ray imaging is the most common radiological test for diagnosing a wide range of lung conditions. Deep Learning (DL) has emerged as a powerful tool for automated analysis of chest X-rays, but it requires large, well-annotated datasets that are often difficult to obtain due to privacy concerns and institutional data restrictions. Federated Learning (FL) addresses this by allowing hospitals to collaboratively train DL models without data sharing. Despite its benefits, achieving robust and balanced generalization across heterogenous data in decentralized settings remains a challenge. We investigate how integrating Adversarial Attacks, which have shown potential in mitigating biases in data skewed regimes, into an FL pipeline impacts both performance and spatial interpretability in chest X-ray analysis. We analyze the viability of an Adversarial Attacks-enhanced FL framework in two simplified scenarios employing public X-ray datasets. Our setup combines the Federated Averaging algorithm with a parameter-efficient ConvNeXt architecture. To enhance both diagnostic accuracy and interpretability in chest X-ray imaging, the network is trained using a targeted single-step adversarial training strategy (Step-ML). We evaluate this framework under controlled, independent and identically distributed conditions, varying both the number of participating clients and the level of data fragmentation, and extend our analysis to a multi-site heterogeneous client scenario by integrating two distinct clinical databases. Our results demonstrate that the Adversarial Attacks-enhanced FL approach improves diagnostic performance compared to both non-adversarial training and non-targeted adversarial baselines. While severe data fragmentation led to an expected decline in absolute performance and spatial interpretability, the aggregated model maintained significantly higher robustness than individual local models. Furthermore, these trends remain consistent across heterogeneous multi-site client regimes. Our findings indicate that combining FL with targeted adversarial training can improve diagnostic performance and interpretability in collaborative chest X-ray analysis while maintaining data privacy. By establishing a controlled baseline, these results constitute a prerequisite for deploying robust deep learning models in complex, real-world clinical environments.