Lightweight Federated Learning for COVID-19, Pneumonia, and TB from Chest X-Ray Images

The COVID-19 virus, often known as the coronavirus, poses a severe threat to international health. The vast amount of chest X-ray (CXR) images may be evaluated by radiologists with deep learning, which could significantly streamline and speed up the diagnosis of pneumonia, TB, and COVID-19. Such methods necessitate extensive training datasets, all of which must be consolidated for processing. Patient information cannot be compiled and shared on a centralized server due to medical data privacy restrictions. This paper describes a federated learning architecture that enables deep learning to be used by different hospitals to screen chest X-rays for COVID-19, Tuberculosis, and Pneumonia without sharing patient data. We investigate the naturally occurring unbalanced data distributions and the fact that they are not independent and non-identically distributed (non-IID) in a federated learning environment. This work offers a federated learning framework that is applicable to different transfer learning models like AlexNet, SqueezeNet, and ResNet architectures. Resnet50 model with a federated learning environment produces 98.73% of classification accuracy results. These results would lead medical institutions to adopt a collaborative process and use the private data rich in COVID-19 and other chest-related diseases to construct a robust model swiftly.