基于超声时频分析与残差网络的生物组织变性识别
Denatured identification of biological tissue based on ultrasonic time-frequency analysis and residual network
刘备 1许克晖 1杨德智 1彭梓齐 1杨江河1
作者信息
- 1. 湖南文理学院数理学院,湖南常德 415000
- 折叠
摘要
为了能对高强度聚焦超声(HIFU)治疗过程中的生物组织进行实时准确的变性识别,提出了一种基于超声时频分析与残差网络(ResNet)的生物组织变性识别方法.首先,采用广义S变换(GST)方法对生物组织超声回波信号进行时频分析,得到二维时频图;然后,通过迁移学习,将在ImageNet数据集上训练得到的参数应用于超声回波信号数据集;最后,利用ResNet101模型从生物组织变性前后的时频图中学习和提取有效的变性信息,并可视化变性特征轨迹,实时地完成生物组织变性识别.实验结果表明:相较于现有基于信号能量,AR系数以及熵特征的变性识别方法,GST-ResNet方法无需人为经验选取特征参数,具有更高的识别率,可以实时准确地完成生物组织的变性识别.
Abstract
In order to achieve real-time and accurate denaturation identification of biological tissues during high-intensity focused ultrasound(HIFU)treatment,a new denatured identification method of biological tissue based on ultrasonic time-frequency analysis and residual network(ResNet)is proposed.Firstly,the generalized S-transformation(GST)method is used to analyze on time-frequency domain of the ultrasonic echo signal and 2D time-frequency image is obtained.Then,the parameters trained on the ImageNet dataset are applied to the ultrasonic echo signal dataset by transfer learning.Finally,the ResNet101 model is employed to learn and extract effective denatured information from the time-frequency map before and after biological tissue denaturation,visualize the characteristic trajectory,and realize the denatured identification of biological tissue in real-time.The experimental results show that compared with the existing denatured recognition methods of AR coefficient and entropy features based on signal energy,GST-ResNet method does not require artificial empirical selection of feature parameters and has higher recognition rate,which can accurately complete denatured identification of biological tissue in real-time.
关键词
时频分析/残差网络/高强度聚焦超声/超声回波信号/变性识别Key words
time-frequency analysis/residual network(ResNet)/high-intensity focused ultrasound(HIFU)/ultrasonic echo signal/denatured identification引用本文复制引用
基金项目
国家自然科学基金资助项目(U2031112)
湖南省自然科学基金资助项目(2023JJ40462)
湖南省教育厅优秀青年项目(22B0694)
出版年
2024
NETL
NSTL
维普
万方数据