«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

33

期数:

2023年09期

页码:

149-154

栏目:

人工智能

出版日期:

2023-09-10

文章信息/Info

Title:

Automated Segmentation of 3D Liver CT Images Using Semi-supervision

文章编号:

1673-629X(2023)09-0149-06

作者:

谢宏彪¹ ; 刘志勤¹ ; 王庆凤¹ ; 黄 俊¹ ; 陈 波¹ ; 周 莹²

1. 西南科技大学 计算机科学与技术学院,四川 绵阳 621010;

2. 绵阳市中心医院,四川 绵阳 621010

Author(s):

XIE Hong-biao1 ; LIU Zhi-qin1 ; WANG Qing-feng1 ; HUANG Jun1 ; CHEN Bo1 ; ZHOU Ying2

1. School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621010,China;

2. Mianyang Central Hospital,Mianyang 621010,China

关键词:

协同训练; 肝脏分割; 半监督学习; 全连接条件随机场; U-Net

Keywords:

cooperative training; liver segmentation; semi-supervised learning; dense conditional random filed; U-Net

分类号:

TP391. 41

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 09. 022

摘要:

肝癌是常见癌症之一,有着较高的死亡率,精准分割肝癌区域是辅助诊断治疗的重要前提。 然而肝脏 CT 图像需要专业的医师进行标注,有标签数据较少,获取途径单一。 针对分割腹部肝脏 CT 图像需要大量高质量标签并且较难获取的问题,提出了一种采用协同训练的半监督学习分割方法。 首先,将有标签数据输入 3D U-Net 和 3D Res U-Net 进行有监督训练,保存训练得到的两个分割模型,在两个分割模型中分别对无标签数据进行预测;然后,挑选预测的结果,再加入全连接条件随机场处理挑选出的伪标签,细化伪标签的边缘信息,提升伪标签的精确度;最后,加入到训练集中,重复上述步骤直到分割结果的 Dice 相似系数停止提升时结束训练。 实验在 LiTS2017 Challenge 肝脏数据集上进行测试,结果表明,在有标签数据集占总数据集的 30% 时,该方法的 Dice 值达到 90. 22% ,几乎与全监督 3D Res U-Net 分割结果持平,说明该半监督学习方法是有效的。

Abstract:

Liver cancer is one of the common cancers with a high mortality rate. Accurate segmentation of liver cancer regions is animportant prerequisite for auxiliary diagnosis and treatment. However,liver CT images need to be labeled by professional physicians,andthere are few labeled data and a single way to obtain them. In order to solve the problem that segmentation of abdominal liver CT imagesrequires a large number of high-quality labels and is difficult to obtain,a semi-supervised automatic segmentation method based on cooperative training is proposed. Firstly,the labeled data is input into 3D U-Net and 3D Res U-Net for supervised training,and the two segmentation models obtained by training are saved. The unlabeled data are predicted respectively in the two segmentation models. Then thepredicted results are selected,and the pseudo - labels selected by the fully connected conditional random field processing are added torefine the edge information of the pseudo-labels. It can improve the accuracy of the pseudo-label,and finally add it to the training set.Repeat the above steps until the Dice similarity coefficient of the segmentation results stops improving to end the training. The experimentwas tested on the LiTS2017 Challenge. The results showed when the labeled data sets account for 30% of the total set,the Dice of theproposed method reaches 90. 22% ,which is almost equal to the fully supervised 3D Res U-Net segmentation result,indicating that thesemi-supervised method is effective.

常用功能

更新日期/Last Update: 2023-09-10