Federated learning with deep convolutional neural networks for the detection of multiple chest diseases using chest x-rays

The increasing global incidence of COVID-19 necessitates the rapid development of a reliable method for diagnosing the disease. The virus is spreading so quickly, that medical personnel are having a difficult time identifying patients who are infected with COVID-19. This is because the symptoms of other chest diseases such as pneumonia, tuberculosis, pneumothorax, lung cancer, and consolidation lung are like COVID-19. Additionally, there is a dearth of testing kits, and it is challenging to ascertain whether the kits provide reliable results. The second problem that occurs is the transfer of data between hospitals located in different parts of the world while adhering to the stringent confidentiality standards imposed by various organizations. When it comes to training a global deep learning model, the major concerns are protecting privacy and constructing a model through collaborative efforts. In this study, we propose a collaborative federated learning (FL) system that makes use of deep learning (DL) models to filter COVID-19 from multiple chest infections (including pneumonia, tuberculosis, pneumothorax, lung cancer, and consolidation lung) using chest X-rays (CXR) obtained from a variety of medical institutions without necessitating the sharing of patient data. We investigate a variety of essential features and components of FL environments, including naturally occurring imbalanced data distributions and non-independent and non-identically distributed (non-IID) data sets, amongst other things. This research study investigates four different DL models, which include Vgg19, DenseNet169, InceptionV3, and DenseNet201. The proposed system experimentally presents that the FL framework achieves better results with these models trained by sharing data. These findings will provide medical institutions with the confidence they need to apply collaborative methods and harness private data to fast construct a credible model for identifying multiple chest diseases.