FedPDN: Personalized Federated Learning with Inter-class Similarity Constraint for Medical Image Classification through Parameter Decoupling

Data heterogeneity across medical centers, resulting in a coupling of universal information for classification tasks and personalized information for private dataset within local models, is still a difficult challenge in Personalized Federated Learning (PFL). Moreover, the high inter-class similarity in the private datasets affects the performance of the local models. Different from pervious works that focus on personalized aggregation or personalized adjusting the global model, we introduce the concept of decoupling universal and personalized information in local models and propose a novel PFL framework for medical image classification in this paper. Specifically, we propose a decoupling strategy at the client side to efficiently utilize universal and personalized information of the local model to solve data heterogeneity. This strategy decouples the parameters of the local models into two components based on Singular Value Decomposition (SVD), namely the universal component and personalized component. The former contains universal information for the classification task, while the later only includes the personalized information for the client dataset. During the training process of PFL, only the universal component is transmitted between the server and clients, which makes our framework has ability to save transmission resource and protect personalized information. To address the challenge of high inter-class similarity in private dataset, during the network training in local clients, we apply an Inter-class Separability (IS) loss to adaptively enlarge the angle between features of different classes in the feature space, thereby reducing the inter-class similarity. Extensive experiments were conducted on a dermoscopic dataset and a glaucoma dataset, achieving accuracy rates of 87.16% and 84.64%, respectively. The results demonstrate that our proposed method outperforms nine advanced methods and achieves state-of-the-art results in the medical image classification tasks.