A Communication Efficient Federated Learning Approach to Multi Chest Diseases Classification

Abstract

Recently, deep learning (DL) has been playing a crucial role in aiding disease detection and started a vital role in medical image analysis. Due to its numerous advantages such as promising accuracy, ease of scalability, and fast diagnosis, DL has been a practical alternative to laboratory-based disease detection methods, especially for the Covid-19 detection. DL also has some disadvantages for privacy concerns, legal regulations, and hardware needs. To address this issue, we adopted the federated learning (FL) approach which is a novel data-private collaborative learning procedure that enables leveraging the big and diverse data securely without collecting the data into a central place. This paper introduces a CNN architecture for multi chest-diseases classification from chest Xray (CXR) images. The proposed classification model has reached % 93.34 accuracy on the task. A combination of various publicly available datasets also presented with 2SS33 CXR images, a mixture of Covid-19, viral or bacterial pneumonia, lung opacity, and normal cases. The proposed FL based model also achieves the tasks with % 92.96 accuracy. Weight pruning and quantization were implemented to reduce the communication cost by 10x of FL which is usually the primary bottleneck in FL. The communication efficient federated training achieves the task with % 92.44 accuracy with a negligible loss in the accuracy. Lastly, the results of central training, federated training and communication efficient federated training were given and compared with the experiments. It is expected that the proposed model might help to encourage organizations or researchers to develop their own models, to improve collaboration, to increase data utility, and to get more benefits from unused data by adopting the federated learning approach.