A federated learning approach for classifying chest diseases from chest X-ray images

A leading global cause of death and morbidity is chest disorders which demand accurate and timely diagnosis for effective management. Chest X-rays serve as a commonly used and widely employed diagnostic tool for detecting these diseases. The evolution of deep learning algorithms in modern years has been essential for early diagnosis and treatment planning by significantly improving the accuracy of classifying chest diseases from chest X-ray images. However, collecting a huge amount of chest X-ray data from multiple healthcare institutions for model training can be challenging due to patient privacy concerns and data regulation laws. Federated learning unfolded as a promising finding to this problem that enables several healthcare institutions to conjointly train a model in balance with patient data decentralized and private. In this study, we propose an approach of federated learning for classifying chest diseases from chest X-ray images employing the NIH Chest X-ray dataset. Our approach involves training the model on the data distributed over multiple healthcare institutions, and aggregating the model weights to obtain a global model. We assess and compare the suggested method with conventional centralized learning techniques using a large-scale chest X-ray dataset. Our outcome shows that federated learning achieves a training and validation accuracy of 80% and 69.86% respectively, while preserving patient privacy and security.