全卷积神经网络与全连接条件随机场中的左心室射血分数精准计算  

Accurate Estimation of Left Ventricle Ejection Fraction Using Fully Convolutional Networks and Fully Connected Conditional Random Field

作　　者：刘晓鸣 雷震[2] 何刊[3] 张惠茅[3] 郭树旭[1] 张歆东[1] 李雪妍[1] Liu Xiaoming;Lei Zhen;He Kan;Zhang Huimao;Guo Shuxu;Zhang Xindong;Li Xueyan(College of Electronic Science and Engineering,Jilin University,Changchun 130012;Center for Biometrics and Security Research,National Laboratory of Pattern Recognition,Institute of Automation,Chinese Academy of Sciences,Beijing 100190;Department of Radiology,The First Hospital of Jilin University,Changchun 130021)

机构地区：[1]吉林大学电子科学与工程学院,长春130012 [2]中国科学院自动化研究所模式识别国家重点实验室生物识别与安全研究中心,北京100190 [3]吉林大学第一医院放射科,长春130021

出　　处：《计算机辅助设计与图形学学报》2019年第3期431-438,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家科技计划项目(2015DFA11180);吉林省自然科学基金学科布局项目(20180101038JC)

摘　　要：左心室射血分数是临床上用于衡量心脏健康的一项重要指标.为提高左心室分割和射血分数计算的精度,提出一种基于改进的全卷积神经网络和全连接条件随机场的方法.首先利用预训练的全卷积神经网络模型对心脏核磁共振影像进行左心室分割并输出概率图;之后采用3D全连接条件随机场对概率图进行后处理,完成像素级的精准密度预测;最后对左心室分割结果进行3D重建,并计算左心室舒张末期容积和收缩末期容积,进而计算出射血分数.实验结果表明,该方法能够实现左心室射血分数的精确且高效的计算,对左心室射血分数的平均预测误差为4.67%,各步骤耗时短.Ejection fraction of left ventricle is regarded as an important metric to measure the status of heart.To enhance the accuracy of left ventricle segmentation and ejection fraction estimation,the paper presents a novel framework which bases on improved fully convolutional networks(FCN)and fully connected conditional random field(fc CRF).Firstly,the framework segmented the region of left ventricle from MRI using a pre-trained FCN and obtained probability maps.Secondly,post-processing of pixel-wise label assignment was performed by 3D fc CRF.Finally,the segmentation was reconstructed in 3D;end-systolic volume and end-diastolic volume were acquired,and ejection fraction of left ventricle were then calculated.The results demonstrate the framework can estimate the left ventricular ejection fraction accurately and efficiently;the mean predicted error of left ventricular ejection fraction is 4.67% and the time-consuming is short.

关 键 词：左心室射血分数计算 深度学习 全卷积神经网络 全连接条件随机场 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]