基于深度学习多尺度特征的胎儿头围测量方法  

FETAL HEAD CIRCUMFERENCE MEASUREMENT METHOD BASED ON DEEP LEARNING MULTI-SCALE FEATURES

作　　者：黄佳伟 付晓薇[1,2] 陈芳 李曦 Huang Jiawei;Fu Xiaowei;Chen Fang;Li Xi(College of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,Hubei,China;Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System,Wuhan 430065,Hubei,China;Department of Ultrasound and Imaging,Wuhan University of Science and Technology Hospital,Wuhan 430065,Hubei,China;School of Artificial Intelligence and Automation,Huazhong University of Science and Technology,Wuhan 430074,Hubei,China)

机构地区：[1]武汉科技大学计算机科学与技术学院,湖北武汉430065 [2]智能信息处理与实时工业系统湖北省重点实验室,湖北武汉430065 [3]武汉科技大学校医院超声影像科,湖北武汉430065 [4]华中科技大学人工智能与自动化学院,湖北武汉430074

出　　处：《计算机应用与软件》2023年第10期222-227,234,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61873323);材料成形与模具技术国家重点实验室开放课题研究基金项目(P2018-016);湖北省自然科学基金项目(2016CFA037);武汉市科技计划项目(2018010401011292);武汉科技大学国防预研基金项目(GF201912);智能信息处理与实时工业系统湖北省重点实验室开放课题项目(znxx2018ZD01)。

摘　　要：传统医学超声胎儿头围测量方法存在效率低、测量不稳定等问题。为此,提出一种医学超声胎儿头围自动测量方法。提出一种H-Unet改进模型进行医学超声胎儿头部图像分割。该过程利用Non-local模块捕获长范围依赖关系,采用空间金字塔模块获取多尺度特征信息,并提出一种改进的损失函数进行网络训练;利用边缘检测算子提取分割目标的头部轮廓;利用基于最小二乘法的椭圆拟合方法自动计算出胎儿头围值。在HC18数据集上进行测试,以头围绝对差值(A_(D))、Hausdorff距离(H_(D))、Dice系数评估该模型性能。实验结果表明:对比已有方法,该方法的上述三种指标均有提升,在胎儿头围测量上可取得满意的效果。Traditional medical ultrasound fetal head circumference measurement methods have problems such as low efficiency and unstable measurement.For this reason,an automatic measurement method is proposed for medical ultrasound fetal head circumference.An improved H-Unet model was proposed for medical ultrasound fetal head images segmentation.During this process,non-local module was used to capture long-range dependencies,the spatial pyramid module was used to obtain multi-scale feature information,and an improved loss function was proposed for network training.The edge detection operator was used to extract the head contour of image segmentation target.The ellipse fitting method based on the least square method was used to automatically calculate the fetal head circumference.Test was conducted on the HC18 data set to evaluate the performance of the proposed model with absolute head circumference(A_D),Hausdorff distance(H_D),and Dice coefficient.The experimental results show that,compared with the existing method,the above indicators of this method are all improved,and satisfactory results can be obtained in the measurement of fetal head circumference.

关 键 词：医学超声图像 空洞卷积 H-Unet网络 多尺度特征 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]