Research On Medical Image Classification Method Based On Horizontal Federated Learning

Medical image is a tool for assisting doctors diagnosed,which is widely used in modern medical diagnosis and treatment.With the advent of artificial intelligence,technologies such as tumor recognition and lesions based on deep learning have emerged.The establishment of these deep learning models requires a lot of data and clearly and accurately marked it.However,in actual life,unlike ordinary image data,medical image data has the following characteristics: The processing and labeling of medical images are strong and have high requirements for accuracy,so the resources are lacking and difficult to obtain;The data contains extremely sensitive sensitive.Privacy information is protected by laws and regulations.Therefore,there is a "data islands" effect between medical institutions;All medical institutions have different data distribution due to their respective attributes and regional environment.The above problems have brought huge challenges to the establishment of medical deep learning models.Therefore,in order to solve the above problems,this article proposes a medical image classification method based on Horizontal Federated Learning,so that all medical institutions can jointly build a higher performance medical image classification model on the premise of protecting patient data privacy.The specific research content is as follows:First of all,in response to the lack of medical image training data and privacy protection,this article introduces Federated Learning,and adds Paillier homomorphic encryption technology to the Federated Averaging Algorithm(Fed Avg)to further protect model parameters security.Then build a Res Net50 and VGG16 models under the overall framework to complete the medical image classification task.Experiments show that on the four public medical image data sets,the accuracy of this method is increased by 3%-8% compared to the model obtained by each medical institution for separate training,which shows that this method has the dual advantage of improving accuracy and privacy protection.Secondly,in response to the problem of decline in model accuracy caused by the data of various medical institutions in non-independent and identically distributed,this article has optimized and improved the local training methods and loss functions in the previous chapter model,and proposed a Personalized Horizontal Federated Medical Image Classification Model(Personalized Horizontal Federated Learning,PHFL).Experiments show that when the four data sets are non-independent and identically distributed(N.I.I.D),the average accuracy of this method(PHFL)has increased by 4%-9%compared to Fed Avg,and it is also better than other methods of solving data N.I.I.D methods in Federated Learning Fed Per and Fed Prox.In summary,Medical Image Classification Method Based on Horizontal Federated Learning in this article can break the bottlenecks encountered by the current medical image classification model,and jointly build a better performance model with many medical institutions under the premise of protecting data privacy.This method can improve the accuracy of classification and help doctors improve diagnosis and treatment efficiency.If you can land to the actual application,it is expected to alleviate the difficulty of seeking a doctor in the current lack of the talents in the current hospital in the grassroots hospital,and ultimately meet the service needs of the masses nearby,convenient,economical,and efficiently seeing medical treatment.