基于不确定性协作图和双层聚合机制的个性化联邦医学图像分割OA

As a distributed machine learning paradigm,personalized federated learning can realize the collaborative training of multi-client models without leaking the original data of the client,and has become a research hotspot in the field of medical image intelligent processing and analysis.However,the existing personalized federated learning methods mainly model client relationships through global collaboration or clustering group collaboration,and their overall collaboration granularity is coarse and lack of flexibility.In recent years,personalized federated learning methods based on collaboration graph model the collaboration relationship between clients by graph structure,which can achieve fine-grained dynamic col-laboration and effectively alleviate the inherent defects of global collaboration and clustering collaboration.However,it on-ly uses the amount of data and model similarity to update the client collaboration graph,and does not consider the inherent high uncertainty in the medical image segmentation task,which makes it vulnerable to high uncertainty clients and reduces the segmentation accuracy.In order to solve this problem,we propose a personalized federated medical image segmentation method based on uncertain collaboration graph and dual-layer aggregation mechanism in this paper.The core advantages of this method mainly include two aspects.Firstly,an uncertainty penalty term is designed and introduced into the server-side objective function to optimize the updating process of the collaboration graph,and generate an uncertain collaboration graph suitable for the medical image segmentation task.By dynamically adjusting the collaboration weights between each client and avoiding knowledge pollution caused by high noise parameters,the stability of collaborative training is effective-ly guaranteed.Secondly,a dual-layer aggregation mechanism based on uncertain collaboration graph is proposed.The first layer of aggregation realizes the local collaboration of clients based on collaboration graph,and mines the effective knowl-edge between similar clients.The second layer of aggregation balances the generality of the global model and the personal-ized requirements of the local client by fusing the local collaborative results and the global model,realizes the effective transfer of high-quality knowledge,and improves the segmentation performance of the client-side local model.In order to fully verify the effectiveness and robustness of the proposed method,a large number of experiments are carried out on four public polyp segmentation datasets.The experimental results show that compared with other advanced medical image seg-mentation methods,the proposed method achieves better segmentation performance on multiple client test data,which pro-vides a new technical solution for personalized federal medical image segmentation in clinical medical scenarios.