超声背散射信号多分支残差网络评估脂肪肝  

Multi-Branch Residual Network for Ultrasonic Classification of Hepatic Steatosis with Backscattered Signals

作　　者：王前 黄勇[1] 吴水才[1] 崔博翔 周著黄[1] WANG Qian;HUANG Yong;WU Shuicai;CUI Boxiang;ZHOU Zhuhuang(Faculty of Environment and Life,Beijing University of Technology,Beijing 100124,China;College of Medicine,Chang Gung University,Taoyuan Taiwan 333323,China)

机构地区：[1]北京工业大学环境与生命学部,北京100124 [2]台湾长庚大学医学院,中国台湾桃园333323

出　　处：《中国医疗设备》2023年第6期77-81,共5页China Medical Devices

基　　金：国家自然科学基金项目(11804013);北京市自然科学基金项目(4222001)。

摘　　要：目的基于超声背散射信号的深度学习技术是超声组织定征中的一种新兴趋势,提出一种基于背散射信号卷积神经网络(Convolutional Neural Network,CNN)的深度学习模型多分支残差网络(Multi-Branch Residual Network,MBRNet),并分析其对脂肪肝的评估效能。方法MBR-Net由3个分支组成,各分支使用不同卷积块,以增强局部特征获取能力,结合带有跳跃连接的残差机制,引导网络有效地捕获特征。实验采用204例临床脂肪肝(S0无脂肪肝80例、S1轻度脂肪肝70例、S2中度脂肪肝36例、S3重度脂肪肝18例)背散射信号。采用背散射信号重构B超图像,手工选定肝实质区域,对其中的每条射频信号,使用长度为768个采样点的门在其上滑动,步长为20个采样点,得到n条门控信号。然后从n条门控信号中随机选取256条。结果共获得261120条信号样本(S0:102400;S1:89600;S2:46080;S3:23040)。MBR-Net与Nguyen网络、Han网络在脂肪肝评估中的性能比较,MBR-Net诊断脂肪肝程度≥S1、≥S2、≥S3均具有更高的准确度、灵敏度和特异性,且MBR-Net的AUC也最高;MBR-Net(三分支网络)的脂肪肝分类效果优于双分支网络和四分支网络。结论相较于传统的单分支、无残差机制的CNN方法,本研究提出的MBR-Net整体上提高了分类精度,在评估肝脏脂肪变性程度的分类任务中取得了良好的性能,MBR-Net可作为超声背散射信号深度学习评估脂肪肝的新方法。Objective The deep learning technology based on ultrasonic backscattered signals is an emerging trend in ultrasonic tissue characterization,to propose a deep learning model multi-branch residual network(MBR-Net)based on backscattered signal convolutional neural network(CNN),and to analyze its evaluation efficiency for fatty liver.Methods MBR-Net was composed of three branches;each branch used different convolutional blocks to enhance the ability of local feature acquisition,combined with the residual mechanism with skip connection,which guided the network to capture features effectively.A total of 204 cases of clinical hepatic steatosis ultrasound backscattered signals(80 cases without hepatic steatosis as S0,70 cases with mild hepatic steatosis as S1,36 cases with moderate hepatic steatosis as S2,and 18 cases with severe hepatic steatosis as S3)were included in the experiments.B-mode ultrasound images were reconstructed by using the backscattered signals,the liver parenchyma area was manually selected.For each of the radio frequency signals,a gate with a length of 768 sampling points was used to slide on it,and the step size was 20 sampling points to obtain n gating signals.Then 256 gating signals were randomly selected from n gating.Results A total of 261120 signal samples were obtained(S0:102400;S1:89600;S2:46080;S3:23040).Compared with Nguyen network and Han network,MBR-Net had higher accuracy,sensitivity and specificity in the diagnosis of fatty liver degree≥S1,≥S2 and≥S3,and the AUC of MBR-Net was the highest.The fatty liver classification effect of MBR-Net(three-branch network)was better than that of two-branch network and four-branch network.Conclusion Compared with the traditional CNN method with single branch and no residual mechanism,the MBR-Net proposed in this study has improved the classification accuracy on the whole and achieved good performance in the classification task of evaluating the degree of hepatic steatodegeneration.MBR-Net can be used as a new method to evaluate hepatic steatosis

关 键 词：超声背散射 深度学习 卷积神经网络 多分支残差网络 脂肪肝 

分 类 号：R318[医药卫生—生物医学工程] TP391[医药卫生—基础医学]