JoWUA

Abstract

In recent years, automated discovery of thoracic diseases from Chest X-ray images (CXR) has become an essential field of computer-aided analysis. The diversity and scarcity of medical data make it complex to generate a precise global classification in the healthcare sector. The main reasons are privacy issues and legal obstacles limiting data sharing between healthcare institutions. On the contrary, data from a single source is hardly enough for developing a universal analysis model. Therefore, Federated Learning (FL) is a potential solution for data diversity and privacy issues. This research uses the FL approach to effectively learn from multiclass and heterogeneous medical data. However, the FL approach is susceptible to adversarial attacks; therefore, blockchain (BC) is integrated to enhance the security and privacy of patient data. This paper proposes the aggregator model, i.e., two-layer Long Short-Term Memory (2LLSTM) with the federated proximal term (FedProx), namely 2LLSTMFP, for enhancing the classification. The proximal term integrated with the 2LLSTM solves the issues created by data heterogeneity and increases stability during classification. 2LLSTMFP aggregates the classified information from the Convolutional Neural Network (CNN). The incorporated FP term develops the utilisation of proximal operators to update 2LLSTMFP weights regularly. The updated 2LLSTMFP weights regularise the effect of adversarial updates with BC, which further improves its robustness to malicious devices, i.e., adversarial attacks. The developed FLBC-2LLSTMFP is evaluated by using the ChestX-ray14 dataset. The FLBC-2LLSTMFP is analysed regarding accuracy, specificity, precision, recall, F1-score, False Positive Rate (FPR), and False Negative Rate (FNR). The existing research MobileLungNetV2 and MLRFNet are used to compare with the FLBC-2LLSTMFP. The accuracy of FLBC-2LLSTMFP is 99.98%, which is higher than that of MobileLungNetV2 and MLRFNet