Dual-Objective Federated Learning Strategy for Lung X-Ray Analysis in Healthcare Systems

Federated Learning has emerged as a promising approach to collaboratively train machine learning models across decentralized healthcare systems with big data while maintaining the privacy of patient data. In this paper, we propose using the joint model to segment and classify lung X-ray images with a federated learning strategy. By eliminating the need for two distinct models, we aim to improve the system’s maintainability while also directly using the output and encoding backbone of the segmentation model in the classifier segment. The presented experiments describe the initialization, training and fine-tuning of the model, with thorough analysis. The results obtained demonstrate the effectiveness of the proposed dual-objective learning strategy, with a segmentation accuracy of 99.15% and 98% mean dice coefficient. The model achieved a score of 94% f1-score in the classification task, considering 4 different possible patient outcomes.