Federated Learning for Thoracic Disease Classification Using Convolutional Neural Networks and Differential Privacy

Early diagnosis of thoracic diseases using chest x-ray imaging remains a critical challenge, particularly in resource-constrained healthcare environments where data sharing is restricted due to privacy concerns. Federated learning (FL) offers a decentralized solution by enabling collaborative model training without sharing sensitive patient data. However, integrating privacy-preserving mechanisms such as differential privacy (DP) introduces additional challenges related to performance degradation and computational overhead. In this study, we present a unified FL framework for multi-label thoracic disease classification using multiple convolutional neural network (CNN) architectures, including ResNet50, DenseNet169, EfficientNet variants and MobileNetV3. Unlike prior studies focusing on single-model evaluation, this work provides a controlled comparative analysis under identical FL settings and investigates the impact of client scalability (5-10 clients) on model performance. Furthermore, we conduct a comprehensive empirical analysis of the privacy utility trade-off by integrating DP with varying privacy budgets (ε = 1, 15 and 30). Experimental results on the CheXpert and NIH Chest x-ray14 datasets demonstrate that the proposed EfficientNet-B3-based federated model achieves a mean AUC of 0.8027, while maintaining robustness across decentralized settings. The integration of DP leads to a predictable reduction in performance, with mean AUC ranging from 0.60 to 0.64, highlighting the inherent trade-off between privacy and diagnostic accuracy. The findings emphasize the practical viability of FL for privacy-sensitive medical imaging applications and provide insights into model selection, scalability and privacy configuration for real-world deployment. The source code for this study is publicly accessible at https://github.com/Zulqarnain8-8/FEDERATED_LEARNING_FOR_THORACIC_DISEASE_CLASSIFICATION.