一种基于U⁃Net网络心脏MR图像左室识别的图割后处理快速算法  

Fast post⁃processing image processing algorithm based on U⁃Net network for left ventricular recognition of cardiac MR images

作　　者：乔艳 苑金辉 谢文鑫 胡晓飞[1] QIAO Yan;YUAN Jinghui;XIE Wenxin;HU Xiaofei(School of Geographic and Bioinformatics,Nanjing University of Posts and Telecommunications,Nanjing 210023,China;College of Telecommunications&Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学地理与生物信息学院,江苏南京210023 [2]南京邮电大学通信与信息工程学院,江苏南京210003

出　　处：《南京邮电大学学报（自然科学版）》2021年第2期29-35,共7页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(61771251);江苏省自然科学基金(BK20171443)资助项目。

摘　　要：为了从心脏MR图像中分割出左心室,提出了一种基于U⁃Net网络的图割后处理算法。先训练U⁃Net网络分割网络得到概率图,接着采用图割算法进行后处理。为了减少图割算法运行时间,采用图像形态学操作去除无需后处理区域,只保留分割结果的边缘区域进行图割优化。采用Dice系数和召回率这两种评价指标,在Sunnybrook数据集上的实验结果表明,该算法结果的边缘更接近标准分割结果,且具有较高的Dice系数、召回率和时间效率。To segment the left ventricle from MR images of the heart,a U⁃Net based graph cutting post⁃processing algorithm is proposed.Firstly,the U⁃Net network segmentation is trained to obtain a probability graph,and then a graph cut algorithm is used for post⁃processing.To reduce the running time of the graph cut algorithm,an image morphology operation is adopted to remove the region without post⁃processing,and only the edge region of the segmentation result is retained for the graph cut optimization.Using the two evaluation index:Dice coefficient and the recall rate,experimental results on the Sunnybrook dataset show that the edge resuls of the algorithm are closer to standard segmentation ones,

关 键 词：左心室分割 心脏MR图像 U⁃Net网络 形态学操作 图割 

分 类 号：TN911.73[电子电信—通信与信息系统]