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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Health

Background:

• Early diagnosis of thoracic diseases via chest X-rays is crucial but hindered by privacy concerns and data sharing restrictions in resource-limited settings.
• Federated learning (FL) offers a decentralized approach for collaborative model training without compromising patient data privacy.
• Integrating differential privacy (DP) into FL frameworks can degrade model performance and increase computational load.

Purpose of the Study:

• To develop and evaluate a unified FL framework for multi-label thoracic disease classification using diverse CNN architectures.
• To comparatively analyze the performance of different CNN models (ResNet50, DenseNet169, EfficientNet, MobileNetV3) within identical FL settings.
• To investigate the impact of client scalability and DP integration on diagnostic accuracy and the privacy-utility trade-off.

Main Methods:

• Implemented a federated learning framework for thoracic disease classification on CheXpert and NIH Chest x-ray14 datasets.
• Evaluated multiple CNN architectures including EfficientNet-B3, ResNet50, DenseNet169, and MobileNetV3 under varying client numbers (5-10 clients).
• Integrated differential privacy (DP) with different privacy budgets (ε = 1, 15, 30) to assess its effect on model performance.

Main Results:

• The EfficientNet-B3 based federated model achieved a mean AUC of 0.8027, demonstrating robustness in decentralized environments.
• Client scalability from 5 to 10 clients showed minimal impact on model performance.
• DP integration resulted in a performance decrease, with mean AUC ranging from 0.60 to 0.64, confirming the privacy-utility trade-off.

Conclusions:

• Federated learning is a viable approach for privacy-sensitive medical imaging analysis, particularly for thoracic disease classification.
• Model selection (e.g., EfficientNet-B3), scalability, and DP configuration are critical factors for real-world deployment.
• The study provides empirical evidence on the performance implications of integrating DP in FL for medical diagnostics.