东北大学学报:自然科学版 ›› 2018, Vol. 39 ›› Issue (11): 1577-1582.DOI: 10.12068/j.issn.1005-3026.2018.11.012

• 信息与控制 • 上一篇    下一篇

基于全卷积网络迁移学习的左心室内膜分割

齐林1, 吕旭阳1, 杨本强2, 徐礼胜1,3   

  1. (1. 东北大学 中荷生物医学与信息工程学院, 辽宁 沈阳110169; 2. 沈阳军区总医院 放射科, 辽宁 沈阳110016; 3. 东北大学 教育部医学影像计算重点实验室, 辽宁 沈阳110169)
  • 收稿日期:2017-08-11 修回日期:2017-08-11 出版日期:2018-11-15 发布日期:2018-11-09
  • 通讯作者: 齐林
  • 作者简介:齐林(1981-),男,吉林长春人,东北大学副教授,博士; 徐礼胜(1975-),男,安徽安庆人,东北大学教授,博士生导师.冯明杰(1971-), 男, 河南禹州人, 东北大学副教授; 王恩刚(1962-), 男, 辽宁沈阳人, 东北大学教授,博士生导师.
  • 基金资助:
    国家自然科学基金资助项目(61773110, 61374015,61202258); 中央高校基本科研业务费专项资金资助项目(N161904002,N130404016,N171904009); 辽宁省博士启动基金资助项目(20170520180).

Segmentation of Left Ventricle Endocardium Based on Transfer Learning of Fully Convolutional Networks

QI Lin1, LYU Xu-yang1, YANG Ben-qiang2, XU Li-sheng1,3   

  1. 1. School of Sino-Dutch Biomedical & Information Engineering, Northeastern University, Shenyang 110169, China; 2. Department of Radiology, General Hospital of Shenyang Military Region, Shenyang 110016, China; 3. Key Laboratory of Medical Image Computing, Ministry of Education, Northeastern University, Shenyang 110169, China.
  • Received:2017-08-11 Revised:2017-08-11 Online:2018-11-15 Published:2018-11-09
  • Contact: XU Li-sheng
  • About author:-
  • Supported by:
    -

摘要: 为了避免过拟合现象,提出了基于全卷积网络迁移学习的左心室内膜分割方法.该方法在已用自然图像训练好的VGGNet模型的基础上对参数进行微调;其次,利用了心室内膜位于MRI图像中心处的先验信息作为选取准则来优化分割结果.将该方法对2009 MICCAI数据集的45个病例进行测试,其DICE指数、APD距离和GC率分别为0.91,1.73mm和97.81%.测试结果表明该方法对于心脏MRI图像的左心室内膜的分割结果较好,当引入一定的先验信息后可以优化测试结果.

关键词: 左心室内膜分割, 深度学习, 全卷积网络, 迁移学习, 核磁共振成像

Abstract: To avoid the over-fitting phenomenon, a segmentation method of left ventricle endocardium based on transfer learning of FCN was proposed. The VGG network which had been trained through the natural images was fine-tuned. In addition, some segmentation criteria were employed to optimizing the results based on the priori information that the left ventricle endocardium was in the center of the MRI(magnetic resonance imaging). In the end, 45 cases taken from the 2009 MICCAI dataset was tested by this mothod. The computed DICE index, APD and GC ratio were 0.91, 1.73mm and 97.81%, respectively. Better results in segmentation of left ventricle endocardium were achieved through the transfer learning of fully convolutional networks and the priori information can improve the automatic segmentation results.

Key words: segmentation of left ventricle endocardium, deep learning, FCN(full convolutional networks), transfer learning, MRI(magnetic resonance imaging)

中图分类号: