基于3D scSE-UNet的肝脏CT图像半监督学习分割方法  

Semi-supervised learning segmentation method of liver CT images based on 3D scSE-UNet

作　　者：刘清清 周志勇[2] 范国华[3] 钱旭升 胡冀苏 陈光强[3] 戴亚康[2,4] LIU Qing-qing;ZHOU Zhi-yong;FAN Guo-hua;QIAN Xu-sheng;HU Ji-su;CHEN Guang-qiang;DAI Ya-kang(School of Biomedical Engineering(Suzhou),Division of Life Sciences and Medicine,University of Science and Technology of China,Suzhou 215163,China;Suzhou Institute of Biomedical Engineering and Technology,Chinese Academy of Science,Suzhou 215163,China;The Second Affiliated Hospital of Suzhou University,Suzhou 215000,China;Jinan Guoke Medical Engineering Technology Development Limited Company,Jinan 250000,China)

机构地区：[1]中国科学技术大学生命科学与医学部生物医学工程学院(苏州),江苏苏州215163 [2]中国科学院苏州生物医学工程技术研究所,江苏苏州215163 [3]苏州大学附属第二医院,江苏苏州215000 [4]济南国科医工科技发展有限公司,山东济南250000

出　　处：《浙江大学学报（工学版）》2021年第11期2033-2044,共12页Journal of Zhejiang University：Engineering Science

基　　金：国家重点研发计划资助项目(2018YFA0703101);中国科学院青年创新促进会资助项目(2021324);苏州市科技计划资助项目(SS201854);丽水市重点研发计划资助项目(2019ZDYF17);泉城5150人才计划资助项目;济南创新团队资助项目(2018GXRC017);江苏省医疗器械联合资金资助项目(SYC2020002).

摘　　要：针对分割神经网络需要大量的高质量标签但较难获取的问题,提出基于3D scSE-UNet的半监督学习分割方法.该方法使用自训练的半监督学习框架,将包含改进的并行空间/特征通道压缩和激励模块(scSE-block+)的3D scSE-UNet作为分割网络.scSE-block+可以从图像空间和特征通道2个方面自动学习图像的有效特征,抑制无用冗余特征,更好地保留图像边缘信息.在自训练过程中加入全连接条件随机场,对分割网络产生的伪标签进行边缘细化,提升伪标签的精确度.在LiTS17 Challenge和SLIVER07数据集上验证所提出方法的有效性.当有标签图像占训练集总图像的30%时,所提方法的Dice相似系数(dice score)为0.941.结果表明,所提出的半监督学习分割方法可以在仅使用少量标注数据的情况下,取得与全监督分割方法相当的分割效果,有效减轻肝脏CT图像分割对专家标注数据的依赖.A semi-supervised learning segmentation method based on 3D scSE-UNet was proposed aiming at the problem that segmentation network requires a large number of high-quality labels and it is difficult to obtain.A self-training semi-supervised learning framework is used and 3D scSE-UNet containing the improved concurrent spatial and channel squeeze and excitation module(scSE-block+)in 3D UNet is utilized as the segmentation network.The scSE-block+can automatically learn effective features of an image from two aspects,image space and feature channel,and suppress redundant features,which helps to preserve more edge information.During the self-training process,dense conditional random field(CRF)is used to refine the segmentation results generated by 3D scSE-UNet,so as to improve the accuracy of the pseudo labels.The effectiveness of the proposed method was verified on LiTS17 Challenge and SLIVER07 dataset.When the labeled images accounted for 30%of the total images in the training set,the dice score of the proposed method was 0.941.Results show that the proposed semi-supervised learning segmentation method can achieve comparable segmentation results with the fully-supervised 3D UNet segmentation method,which effectively reduces the dependence on expert labeled data in liver CT images segmentation.

关 键 词：半监督学习 自训练 3D UNet 注意力模块 全连接条件随机场 

分 类 号：R318.14[医药卫生—生物医学工程]