深度学习在泌尿生殖裂孔面积全自动测量中的应用价值  

Application Value of Deep Learning in Fully-automatic Measurement of Urogenital Hiatus

作　　者：曾丽娟[1] 罗欢嘉 杨鑫 林泽慧 倪东 温红[1] 陆喜多 Zeng Lijuan;Luo Huanjia;Yang Xin;Lin Zehui;Ni Dong;Wen Hong;Lu Xiduo(Huizhou Central People's Hospital,Huizhou,Guangdong 516001.China;National Regional Key Technology Engineering Laboratory for Medical Ultrasound Guangdong Provincial Key Laboratory of Biomedical Measurements and Ultrasound Imaging School of Biomedical Engineering,Health Science Center,Shenzhen University,Shenzhen.Guangdong 518060,China)

机构地区：[1]惠州市中心人民医院,广东省惠州市516001 [2]医学超声关键技术国家工程实验室,广东省生物医学信息检测与超声成像重点实验室,深圳大学医学部生物医学工程学院,广东省深圳市518060

出　　处：《中国超声医学杂志》2021年第12期1408-1412,共5页Chinese Journal of Ultrasound in Medicine

基　　金：惠州市科技计划项目(No.210331214571992)。

摘　　要：目的探讨基于深度学习的卷积神经网络(CNN)在经会阴盆底超声泌尿生殖裂孔面积(UH)全自动测量中的应用价值。方法采集1 200幅UH最小裂孔三维图像作为训练集,在GE、Philips超声机上采集160幅UH图像作为测试集。首先,由3位经验丰富的超声医师在训练集上进行标记,用于训练CNN模型,获取全自动测量软件,在测试集上测量UH参数;再由2位不同年资的盆底超声医师(D1、D2)在测试集上分别进行手动测量,测量参数包括UH面积、周长、前后径、左右径、右侧肛提肌尿道间隙、左侧肛提肌尿道间隙,对测量耗时以及UH结果进行记录并进行统计学分析。结果 GE组中,全自动测量和D1、D2手动测量的组内相关系数(ICC)分别为0.738~0.983,0.730~0.959,均P<0.01,Pearson相关系数r分别为0.749~0.969,0.755~0.962,均P<0.01;Philips组中,全自动测量和D1、D2手动测量的ICC值分别为0.693~0.958,0.515~0.956,均P<0.01,Pearson相关系数r分别为0.781~0.966,0.546~0.957,均P<0.01;全自动测量两次测量间的ICC值为1;全自动测量平均耗时<0.2 s。结论基于CNN构建的全自动测量软件对不同超声机器的数据表现出可靠的可信度,极大提高了UH的测量效率,在盆底超声检查中有较高的应用价值。Objective To investigate the application value of convolutional neural network(CNN) based deep learning techniques in fully-automatic measurement of urogenital hiatus(UH) by transperineal pelvic ultrasound. Methods 1200 three-dimensional images of the UH with minimal hiatus were acquired and annotated by 3 experienced experts for the CNN model training. 160 UH images were acquired on GE、Philips ultrasound instruments for testing. CNN model provides automated measurements for the UH images in the testing set. Then, two pelvic ultrasound doctors(D1, D2) with different experience levels performed manual measurement for the testing set. The measurement parameters of UH includes area, circumference, anterior-posterior diameter, left-right diameter, left and right levator urethral gap distance. The measurement results of UH and the time cost were recorded respectively for statistical analysis. Results Differences were compared between the fully-automatic results and the manual measurements from D1 and D2. In the GE group, the intra-class correlation coefficients(ICC) got the ranges of 0.738-0.983, 0.730-0.959(all P<0.01) and the Pearson correlation coefficients achieved the ranges of 0.749-0.969, 0.755-0.962(all P<0.01);In the Philips group, ICC values presented the range of 0.693-0.958, 0.515-0.956(all P<0.01) and the Pearson correlation coefficients got the range of 0.781-0.966, 0.546-0.957(all P<0.01). The ICC value for the multiple fully-automatic measurements was 1;the mean time cost for the automated measurement was less than 0.2 s per image. Conclusions The CNN based fully-automatic measurement software shows high reliability to the ultrasound images from different machines, which greatly improves the efficiency of UH measurement and has important application value in the routine pelvic ultrasound examinations.

关 键 词：盆底超声 泌尿生殖裂孔 深度学习 全自动测量 

分 类 号：R445.1[医药卫生—影像医学与核医学] R711[医药卫生—诊断学]