Design And Implementation Of Medical Image Segmentation System Based On Federated Learning

Federated Learning is a distributed machine learning approach that involves training local models on individual devices and aggregating the model parameters to create a global model,enabling collaborative learning while preserving privacy.However,due to the disparate data distributions on local devices,the performance of the global model may vary across different devices.This study primarily focuses on model aggregation algorithms and model updating algorithms under non-iid(non-independent and identically distributed)data.To address the limitations of existing aggregation algorithms in non-iid settings,this study proposes a heuristic aggregation algorithm called FedCsf.The algorithm improves aggregation performance in non-iid scenarios by considering the performance of each client model across different classes and their data distribution characteristics.Experimental results demonstrate that the proposed algorithm outperforms other aggregation algorithms,particularly on the CIFAR-10 image recognition dataset.In federated learning,client data exhibits non-iid characteristics,and the aggregation of the global model often leads to the loss of personalized advantages.Therefore,this study introduces a local model updating algorithm called LocalAvg,which achieves personalization by performing weighted averaging between the local and global models.Experimental results show that the LocalAvg algorithm achieves the highest accuracy on the EMNIST and CIFAR-10 datasets,outperforming other updating algorithms.A healthcare image segmentation system based on federated learning is designed and implemented.The system consists of an administrator module,a doctor module,and a general module.The system adopts a RESTful API design and utilizes the HTTP protocol for communication between clients and servers.The V-Net model is employed for image segmentation,and the federated learning process is used for model aggregation and updating.The front-end presentation page is developed using the Vue.js framework,while the back-end information processing utilizes Spring Boot.The system undergoes performance testing with Apache JMeter to ensure stable interfaces and responsive behavior.With effective feature design and technology selection,the system achieves promising practical results.