Abstract
In light of the COVID-19 pandemic, the need for a chest X-ray scans classifier is crucial in order to diagnose patients and classify scans into normal, COVID-infected, and pneumonia. Federated learning was chosen for the classification as it uses a decentralized approach to train the model at the local servers belonging to each entity in various geographic locations. Therefore, information leakage that could happen from the traditional centralized approach of training is prevented, besides saving the huge cost of central storage. However, between the vast difference in the number of X-ray scans per data-silo (i.e. hospital), the dissimilar image-acquisition techniques, and the diverse morphological structures of the human chest, non-IID (non-Independent and Identically Distributed) skews are introduced in the data. In this paper, real-world datasets of COVID and pneumonia scans are used to satisfy all the non-IID data skews. An experiment was then conducted to test the effect of these skews using five federated learning algorithms, FedAvg, FedProx, FedNova, SCAFFOLD, and FedBN, under the same metrics. The obtained accuracy values are 79.5%, 76.92%, 5.57%, 79.18%, and 84.4%, respectively. In this paper, we present the different effects of non-IID skews on the training process and discuss the different federated learning variations to mitigate the data heterogeneity.