基于全卷积神经网络的左心室图像分割方法  

Left Ventricular Image Segmentation Method Based on Full Convolutional Neural Network

作　　者：谢文鑫 苑金辉 胡晓飞[2] XIE Wen-xin;YUAN Jin-hui;HU Xiao-fei(School of Telecommunication&Information Engineering,Nanjing University of Posts and Telecommunications;School of Geography&Bioinformatics,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

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

出　　处：《软件导刊》2020年第5期19-22,共4页Software Guide

基　　金：国家自然科学基金项目(61271082);江苏省自然科学基金项目(BK20141432);江苏省重点研发计划项目(BE2015700)。

摘　　要：针对左心室在心脏图像中面积较小,且存在样本数量不平衡等问题,将一种基于Tversky系数的损失函数应用于心脏左心室分割模型训练。在分割模型中加入注意力模块,当低层特征向高层特征传递图像语义信息时,抑制低层特征图中与分割目标不相关区域,减少这些区域对分割结果的干扰。将以上两种方法结合应用到多输入多输出的全卷积神经网络中,获得心脏左心室图像分割结果。实验结果表明,改进后的算法在原有基础上Dice系数提高了3.3%,召回率提高了4.8%。Aiming at the problem of imbalance in the number of samples due to the small area of the left ventricle in the cardiac image,we use the loss function based on Tversky coefficient to train the left ventricular segmentation model of the heart.Secondly,the atten⁃tion module is added to the segmentation model.When the low-level features transfer the image semantic information to the advanced features,the regions in the low-level feature map that are not related to the segmentation target are suppressed,and the interference of the regions on the segmentation results is reduced.The above two methods are combined into a multi-input and multi-output full convo⁃lutional neural network to obtain the segmentation result of the left ventricular image of the heart.The experimental results show that the improved algorithm increases the Dice coefficient by 3.3%and the recall rate by 4.8%.

关 键 词：全卷积神经网络 图像分割 损失函数 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]