Unsupervised Medical Image Segmentation Algorithm Based On Generative Adversarial Network

With the development of big data and computer computing power,deep neural networks have achieved great success in recent years,but training of deep neural network models requires a large amount of labeled data.Thus unsupervised domain adaptation techniques have received more and more attention.The purpose of domain adaptation is to alleviate the problem of needing to retrain pre-trained models when they are applied to different domains,as it requires a large amount of additional training data from the target domain.Especially in the medical area,annotating the data requires expert knowledge and can be costly in terms of labor.In this paper,we propose an innovative Dual-Scheme Fusion Network(DSFN)for this problem,which performs co-alignment of source and target domains from both image level and feature level perspectives for unsupervised domain adaptation in medical image segmentation problems.In particular,this paper utilizes adversarial learning and deeply supervised mechanisms at multiple levels to perform cross-domain translation on the appearance of images while extracting domain-invariant features.The network builds the connection of two image-level adaptation schemes from the source domain to the target domain and from the target domain to the source domain through a cycle architecture.This balanced joint information flow helps to reduce the domain gap,and fusing the results of these two schemes can further improve the performance of the network.Also,this paper performs the tasks of image segmentation and image generation simultaneously in the form of a joint network,where both tasks share a feature encoder.This feature encoder is involved in both image-level and feature-level adaptation,and their complementary advantages are exploited through end-to-end learning.Experiments show that the method proposed in this paper can achieve significant performance improvements compared to other state-of-the-art domain adaptation methods.Ablation experiments are also conducted to analyze each component in the network.