FedSSH: A Consumer-Oriented Federated Semi-Supervised Heterogeneous IoMT Framework

Abstract

Medical image analysis is essential for disease diagnosis and treatment. However, its advancement is hindered by challenges such as data privacy concerns, heterogeneity in computational resources, and the scarcity of labeled data. To address these issues, we propose FedSSH, a four-layer Internet of Medical Things (IoMT) federated semi-supervised heterogeneous framework, comprising the data, end, edge, and cloud layers. FedSSH facilitates privacy-preserving and efficient cross-layer collaborative computation. By integrating model heterogeneous federated learning (MHFL), it enables devices with varying computational capacities to train models of different complexities, thereby enhancing resource utilization and empowering consumers with access to intelligent services on personal devices. Furthermore, the incorporation of semi-supervised learning allows the model to effectively leverage unlabeled medical images, thereby improving generalization performance. Experimental results on multiple datasets, including CIFAR-10 and MedMNIST, demonstrate that FedSSH surpasses traditional federated learning (FL) methods (FedAvg, FedProx), demonstrating its adaptability to heterogeneous data distributions. Additionally, even with limited labeled data, FedSSH achieves performance comparable to fully supervised training, highlighting its ability to leverage unlabeled data efficiently. Convergence analysis further validates the advantages of cloud-edge collaborative optimization in accelerating training and stabilizing model performance. FedSSH offers an efficient and privacy-aware FL framework for medical AI, with the potential to drive intelligent medical image analysis towards consumer applications.

Topics & Concepts