Federated Performance-based Averaging (FedPA): A Robust and Selective Learning Framework for Chest X-Ray Classification in Heterogeneous Data Environments

Chest X-ray imaging remains a cornerstone in the diagnosis of thoracic conditions such as COVID-19, pneumonia, and lung opacity. Despite advancements in deep learning, the development of robust and generalizable models is limited by data privacy constraints, as patient data cannot be centralized across institutions. Federated Learning (FL) has emerged as a promising solution by enabling collaborative model training without sharing raw data. However, standard FL algorithms like FedAvg, FedProx, and FedSGD aggregate all client updates without considering their individual quality, making them vulnerable to performance degradation in the presence of data heterogeneity, label noise, or underperforming clients. To address these challenges, this study proposes Federated Performance-Based Averaging (FedPA), a novel selective aggregation strategy that incorporates only those client models that meet a pre-defined performance threshold during training. By leveraging an accuracy-based filtering mechanism, FedPA ensures that only sufficiently trained and reliable local models contribute to global updates. The method was evaluated on a multi-class, non-IID chest X-ray dataset containing four classes: Normal, COVID-19, Pneumonia, and Lung Opacity. Using DenseNet as the backbone model, experiments were conducted across four federated clients, each biased toward a specific class to simulate real-world data distributions. Results demonstrate that FedPA significantly outperforms baseline federated algorithms across key metrics, achieving a global accuracy of 91.82%, F1-score of 92.48%, and recall of 92.08%. The method also achieved faster convergence, higher stability, and reduced round-to-round accuracy fluctuations. System-level evaluations further show that FedPA offers competitive efficiency in terms of inference time, throughput, CPU usage, and memory footprint, making it suitable for deployment in resource-constrained clinical environments. Overall, FedPA offers a practical and effective advancement in federated learning for medical imaging. By filtering unreliable client contributions, it preserves model quality and privacy, presenting a viable path for clinical deployment in scenarios where data centralization is infeasible due to ethical, legal, or logistical constraints.