面向心脏MRI分割的半监督空间一致性约束网络  

S^3CNet:Semi-Supervised Spatial-Consistency Constrained Networks for Cardiac MRI Segmentation

作　　者：李才子 刘瑞强 司伟鑫[1] 袁志勇[3] 王平安[1,4] Li Caizi;Liu Ruiqiang;Si Weixin;Yuan Zhiyong;Heng Pheng Ann(Shenzhen Institutes of Advanced Technology,Chinese Academy of Sciences,Shenzhen 518055;University of Chinese Academy of Sciences,Beijing 100049;School of Computer Science,Wuhan University,Wuhan 430072;Department of Computer Science and Engineering,The Chinese University of Hong Kong,Hong Kong 999077)

机构地区：[1]中国科学院深圳先进技术研究院,深圳518055 [2]中国科学院大学,北京100049 [3]武汉大学计算机学院,武汉430072 [4]香港中文大学计算机科学与工程学系,香港999077

出　　处：《计算机辅助设计与图形学学报》2020年第7期1145-1153,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(U1813204,61802385).

摘　　要：精准分割心脏磁共振图像(MRI)分割对于心脏功能分析至关重要.当前基于数据驱动的神经网络模型极大地促进了心脏MRI分割的发展,然而对标注数据的依赖极大地限制了神经网络模型在心脏MRI分割领域的应用.为了降低神经网络模型对于标注数据的依赖,提出一种基于无监督空间一致性约束的半监督心脏MRI分割方法,在少量有标注数据的监督学习基础上,利用无标签数据在模型输入端和输出端分别进行空间变换后前后一致的特性,构建对于无标注数据的空间一致性约束.使用ACDC 2017心脏多组织分割数据集评估了所提出的方法,实验结果表明,相对于有监督学习,通过无监督数据的空间一致性约束能够显著提升模型的泛化能力;此外,相对于其他state-of-the-art的半监督方法,文中方法也拥有更优的泛化性能.Precise cardiac MRI segmentation is crucial for cardiac function analysis.Currently data-driven neural network greatly promotes the development of cardiac MRI segmentation,whereas the dependence on annotated data has limited the generalization of deep neural networks.For the purpose of reducing the dependence on annotated data for deep neural networks,we propose a semi-supervised spatial-consistency networks(S3 CNet)for cardiac MRI segmentation,which employs the spatial transformation consistency of unannotated data before and after being fed into a network architecture.We evaluate the performance of our proposed method on ACDC 2017 cardiac multi-structural segmentation dataset,demonstrating that our method can achieve much better generalization performance than that with supervised learning,meanwhile,the generalization of the proposed method is also better than other state-of-the-art semi-supervised methods.

关 键 词：心脏MRI分割 空间一致性约束 半监督学习 无监督数据增强 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]