Research On Domain Adaptation Technology In Brain Disease Lesion Segmentation Scence

In recent years,with the rapid development of machine learning technology,target detection,semantic segmentation and other fields have gradually emerged a lot of mature technical solutions.These machine learning methods also provide new research ideas for intelligent analysis of neuroimaging.However,traditional deep learning methods usually rely on massive data,and require that the data generation mechanism does not change with the environment.However,such assumptions are often too harsh for medical imaging applications.One is that the number of patients that each medical center can receive is limited,and it is very difficult to obtain massive data.Second,the supervision label of medical data must rely on experienced radiologists,which makes pixel level labeling more expensive and difficult to obtain.How to analyze and mine large-scale data in non-stationary environment is one of the most challenging frontier directions of modern machine learning.Transfer learning relaxes the constraint that training data and test data must be independent and identically distributed in traditional machine learning,so it can mine domain invariant essential features and structures between two different but interrelated domains,so that domain invariant knowledge between different domains can be transferred and reused between domains.Transfer learning is the basic method to solve the problem of the scarcity of target task annotation data,and its research is still in a challenging stage.It is found that there are two knowledge bases worth transferring in the task of intelligent detection of brain medical images.First,for a certain disease,although the labeled data is scarce,the unlabeled data is very easy to obtain.In these unlabeled data,there are abundant general visual expression of brain tissue structure.Using unsupervised and semi supervised methods to fully mine these knowledge and transfer knowledge can provide good prior knowledge for supervised training.Second,for a single brain disease,there is insufficient labeled data,but other brain disease labeled data can be obtained.These other brain diseases,with the same background and abnormal signals caused by lesions,can provide knowledge supplement for the current study of diseases.In this project,we first focus on the network framework of deep learning and transfer learning,and propose domain adaptation solutions for brain disease focus segmentation from two application directions of unmarked data of the same disease and labeled data of other brain diseases.For unlabeled data of the same disease,considering that neural network is a representation learning method,this paper proposes to use auxiliary task to mine the general visual features of brain environment,and realize the supervision of feature extractor through comparative learning.For the labeled data of different diseases,due to the different distribution of different data sets,it is easy to have negative migration and under adaptation problems in the process of migration.This paper proposes a domain adaptation method of online dynamic weight allocation,which effectively avoids the phenomenon of under adaptation and negative transfer in knowledge transfer.This paper proves that the unlabeled data of the same disease and the labeled data of different diseases are valuable for brain diseases.In view of these two application fields,this paper puts forward the corresponding solution and verifies its effectiveness through experiments.These schemes have good expansibility and can provide valuable reference for medical image transfer learning.