Collaborative Segmentation of COVID-19 From non-IID Topographies in the Internet of Medical Things (IoMT)

Abstract

The Internet of Medical Things (IoMT) offers numerous advantages in the diagnosis, monitoring, and treatment of a wide variety of illnesses for both patients. COVID-19 has caused a global pandemic and turned out to be the utmost crucial danger threatening the whole world. Thus, scholars’ attention moved toward Deep learning (DL) and IoMT for developing automated systems for COVID-19 diagnosis andor prognosis based on chest computed tomography (CT) scans, and it has shown great success in several tasks, including classification and segmentation. Nevertheless, developing and training a superior DL approach necessitates accumulating a substantial amount of patients’ CT scans together with their labels. This is an expensive and time-consuming task that restricts attaining large enough data from a single siteinstitution, However, owing to the necessity for protecting data privacy, it is difficult to accumulate the data from several sites and store them at a centralized server. Federated learning (FL) alleviates the need for centralized data by spreading the public segmentation model to different institutional models, training the segmentation model at the institution, and followingly calculating the mean of the parameters in the public model. Nevertheless, researchers advocated that private information could be restored using the parameters of the model. This study presents a privacy-protection technique for the challenge of multi-site COVID-19 segmentation. To tackle the challenge, we introduce the FL technique, in which a distributed optimization procedure is developed, and randomization techniques are proposed to change the joint parameters of private institutional segmentation models. Bearing in mind the complete heterogeneity of COVID-19 distributions from diverse institutions, we develop two domain adaptation (DA) techniques in the proposed FL design. We explore several applied characteristics of optimizing the FL approach and analyze the FL approach in comparison with alternate training approaches. Finally, the results validate that it is auspicious to employ multi-site non-shared CT scans to improve the COVID-19 infection segmentation.