Research And Application Of Knee Joint Image Diagnosis Algorithm Based On Federated Learning

With the emphasis on data privacy,it becomes increasingly difficult to collect data sets from different parties for model training.Therefore,in order to solve this "data island" problem,federated learning,as a distributed training framework,no longer collects all data sets together and trains the model,but conducts collaborative training through parameter sharing.However,domain shift may occur when data sets of different participants are used under this framework.Most of the previous relevant works adopted global models with a single structure,which often has certain limitations and unreliable in training.Meanwhile,average method was simply adopted in global model aggregation,without considering the impact of data sets differentiation on model aggregation,which limits the final performance of the model.At the same time,the problem of "data island" is more prominent in the field of medical image analysis due to data sensitivity,and there are few end-to-end federated learning methods applied in this field.Therefore,we propose a new federated learning method for the above problems,and apply it for the first time to the knee injury prediction problem,which is widely studied in medical image analysis.The knee meniscus injury accounts for about50% of the causes of knee injury,and it occurs in different ages.In clinical diagnosis,for most patients,doctors often use magnetic resonance imaging(MRI)for preoperative evaluation,compared with the invasive method of arthroscopy.However,the site of meniscus injury accounts for a relatively small proportion in knee MRI,the diagnosis process is time-consuming and labor-intensive,and the type of injury is complex,which makes it difficult for the junior imaging doctors to diagnose.Therefore,this paper makes the following research on knee MRI diagnosis algorithm and application based on federated learning.(1)This paper proposes a weighted federated learning method based on Mixture of Experts domain adaptation.The model uses federated learning to solve the "data island" problem of knee MRI,and when facing the problem of domain shift,it introduces a variety of public models with different structures to share parameters,and makes a final prediction through the Mixture of Experts layer dynamically combining the output of private models and public models with different structures.At the same time,when aggregating the public models,by measuring the similarity between different data sets,the corresponding weight matrix is generated to guide the parameter aggregation of the public models,and a unique public model is generated for each participant.We apply the proposed method to the multi-site Magnetic resonance imaging end-to-end classification,and the experiments demonstrate its effectiveness.(2)This paper designs and implements the knee joint imaging diagnosis system.The system embeds a pre-training model based on the innovative method proposed in this paper,which is used to predict whether the imported knee MRI is injured or not,and supports users to simulate the training process of federated learning locally.In addition,according to the actual use requirements of users,the system also developed the following functions,including automatic meniscus segmentation,MRI viewing,data preprocessing,patient basic information and historical diagnosis information management.These functions are convenient for users to manage various information of patients,and assist users to make more accurate clinical diagnosis.