基于深度协同训练的肝脏CT图像自动分割方法  

Automatic liver segmentation from CT images based on deep co⁃training

作　　者：姜威威 刘祥强 韩金仓 JIANG Wei⁃wei;LIU Xiang⁃qiang;HAN Jin⁃cang(School of Information Engineering,Lanzhou University of Finance and Economics,Lanzhou 730020,China)

机构地区：[1]兰州财经大学信息工程学院,甘肃兰州730020

出　　处：《电子设计工程》2020年第14期175-179,共5页Electronic Design Engineering

摘　　要：文中提出了一种肝脏在CT(Computed Tomography)图像中的半监督自动分割方法。该方法采用深度协同训练模型以解决医学图像领域中有标签数据获取困难且成本高的问题。首先利用有标签数据建立U-Net和2D V-Net两种分割网络,并分别对无标签数据进行分割,然后对分割结果进行粗略挑选,再进行精细挑选,最后将置信度较高的伪标签加入到训练集中,重复此过程直到对验证集分割结果的Dice值不再增大时为止。提出的方法可以减少迭代过程中累积的误差,在2017 Liver Tumor Segmentation(LiTS)数据集上的结果表明,该方法与全监督学习相比可以有效提高分割精度。This paper presents a semi⁃supervised automatic liver segmentation method for Computed Tomography(CT)images.The method relies on deep co⁃training to address the problem that labeled data are difficult and costly to be obtained in medical image field.Two segmentation networks,U⁃Net and 2D V⁃Net,are built by the labeled data,and then they provide the segmentation results of the unlabeled data separately.The segmentation results are roughly selected first,and then followed a carefully selection.Finally,the pseudo⁃labels with high confidence are added to the training set.This procedure is repeated until the Dice value does not increase any more.By doing so,our method can reduce the accumulated error.The experimental results on the 2017 Liver Tumor Segmentation(LiTS)dataset show that our method can improve the segmentation accuracy compared with the full supervised learning methods.

关 键 词：肝脏自动分割 CT图像 半监督学习 分割网络 深度协同训练 

分 类 号：TN957.52[电子电信—信号与信息处理]