A Distributed Federated Learning Platform For Medical Image Segmentation

Medical image segmentation plays a key role in computer-aided diagnosis.But due to privacy constraints,it is always difficult to directly aggregate medical image data from various institutions to train a highly precise model.Federated learning provides a privacypreserving method that allows not to aggregate data from each client,greatly improving security.So it has attracted attention in the field of medical image segmentation in recent years.Most of the existing federated learning methods assume that the client has complete labeled data.However,in the field of medical images,due to the high cost of labeling and the subjectivity of the labeling standard,some medical institutions don’t have too much high-quality labeled data although they have a large amount of data,and a large amount of unlabeled data is available locally.In addition,due to the difference of clinical settings,patient groups and devices used,image data from different sources is often heterogeneous,and a single global model would be difficult to perform well on heterogeneous data from various institutions.To address the above issues,the main work and innovation points of this thesis are as follows.1.For the utilization of unlabeled data from each client,this thesis proposes a federated semi-supervised approach based on feature consistency and pseudo-labeling.The method uses consistency and pseudo-labeling to exploit unlabeled data,and combines federated learning to implement the federated semi-supervised method and use regularization terms to prevent local drift.we finally achieve the full utilization of valid data from each federated learning client,including labeled and unlabeled data,to improve the performance of federated model.2.This thesis proposes a personalized federation learning method based on personalized layer and dynamic weights to solve the problem of global model deterioration due to data heterogeneity by allowing clients to learn personalized models based on their own data distribution through personalization layer,and to reduce the negative impact of local model differences on the global model through dynamic weights.3.This thesis implements a distributed federated learning platform for medical image analysis based on the FATE,an open source federated learning framework.This platform integrates the proposed semi-supervised segmentation model and personalized federated learning methods,and provides the ability to customize indicators and analyze the results.