Privacy-Preserving Federated Deep Learning for Medical Image Segmentation in IoT-Enabled Healthcare

Abstract

The increasing use of Internet of Things (IoT) technologies in healthcare has enabled a great deal of medical data to be produced and used by medical imaging devices and institutions in the industry. Deep learning has shown great promise as a tool for automated medical imaging segmentation. However, regular centralized learning systems require images to be sent to one specific server that creates potential privacy, security, communication, and legal issues. Federated learning (FL) offers a solution by allowing separate medical institutions to train a deep learning model cooperatively without actually exchanging their raw medical data. Nevertheless, medical image segmentation is still a challenge due to the fact that such institutions have heterogeneous and non-independent and identically distributed (non-IID) data properties, while all the IoT devices have limited computational and resource possibilities. In this paper, a privacy-preserving federated deep learning structure for medical image segmentation in IoT-supported healthcare setting is proposed. The mentioned approach includes a U-Net-based segmentation technology, while in the process of federated training differential privacy is utilized to prevent possible information leakage from the updates. A special attention is given to resource-aware federated aggregation methods that optimize learning in heterogeneous IoT surroundings. The performance of the mentioned framework is evaluated by means of segmentation accuracy, Dice coefficient, intersection-over-union (IoU), precision, recall, communication costs, and time used for training. The experimental environment of the study is a simulated multi-client healthcare setting where medical imaging data is distributed between geographically different from each other clients that deal with the problem of heterogeneous data distribution. The results of the experiments showed that the proposed technique was able to achieve Dice coefficient close to 0.89 value, IoU coefficient of 0.81, as well as to lower communication costs as compared with the conventional federated learning approach.

Keywords: Federated learning, medical image segmentation, Internet of Things, healthcare IoT, deep learning, differential privacy, U-Net, privacy preservation, edge computing.