摘要
胎儿标准丘脑水平横切面是胎儿双顶径与头围的测量切面,而双顶径和头围这两个测量参数对于预测胎儿体重有重要的作用。临床上此切面一直由超声医生手动获取,手动获取的切面质量高度依赖超声医生的临床工作经验,不但耗时,而且容易得到图像质量较差的切面。为了解决手动获取存在的问题,提出一种基于更快速的区域卷积神经网络(faster R-CNN)的胎儿头围超声图像质量控制方法,辅助医生自动、快速和准确地获得标准丘脑水平横切面。首先,与超声专家团队制定评定协议,通过数据增强的方法,构建胎儿头围超声图像数据库;然后,通过faster R-CNN从训练数据中学习提取有识别性的特征,并利用通过联合训练和交替优化,使得区域建议网络(RPN)模块和fast R-CNN模块共享卷积层特征,构建一个完全端到端的卷积神经网络(CNN)对象检测模型,检测关键解剖结构;最后,通过检测的解剖结构结果对丘脑水平切面进行自动评分,根据评分结果进而自动判断是否是标准切面。对所采集的513张超声切面,80%的作为训练数据集,20%为测试数据集。所提出的方法能够准确地定位到丘脑水平横切面的5个解剖结构,5个解剖结构的检测平均准确度达到80.7%,且每张丘脑水平切面的检查时间大约0.27s。所提出的方法对胎儿头围超声图象进行自动化质量控制是可行的。
The transthalamic plane of fetal is used to measure the biparietal diameter and head circumference of the fetus,and these two measurement parameters play an important role in predicting the fetal weight.Clinically,the plane has always been manually acquired by the ultrasound doctor,and the quality of the manually obtained plane is highly dependent on the clinical work experience of the doctor,which is time consuming,and the poor image quality for plane often happen.In order to overcome the problem of manual acquisition,we proposed a novel method for the quality assessment of fetal head ultrasound images based on faster region-based convolutional neural networks(faster R-CNN),aiming to help doctors automatically,quickly and accurately obtain the standard transthalamic plane.Frist,we set up an evaluation protocol with the team of ultrasound experts and build a database of fetal head ultrasound images through data-enhanced methods.Second,faster R-CNN could learn from the training data to extract useful features,and through the use of joint training and alternative optimization,so that the regional proposal networks(RPN)module and fast R-CNN module shared the convolution layer features and built a complete end-to-end CNN object detection model to detect the key anatomical structures.Finally,the transthalamic plane was automatically scored by the results of the detected anatomical structure,and according to the score results,it was automatically determined whether the plane was a standard one.We collected 513 ultrasound planes,80% was used as a training dataset and 20%as a test dataset.Our method could accurately locate the five anatomical structures of the transthalamic plane with an average accuracy of 80.7% and the examination time of each ultrasound image was approximately 0.27 s,which indicated that it was feasible to perform automated quality control of fetal head ultrasound images by the proposed method.
作者
林泽慧
雷柏英
姜峰
倪东
陈思平
李胜利
汪天富
Lin Zehui;Lei Baiying;Jiang Feng;Ni Dong;Chen Siping;Li Shengli;Wang Tianfu(School of Biomedical Engineering,Shenzhen University,Guangdong Key Laboratory for Biomedical Measurements and Ultrasound Imaging,Shenzhen 518060,Guangdong,China;Department of Ultrasound,Affiliated Shenzhen Maternal and Child Healthcare,Hospital of Nanfang Medical University,Shenzhen 518060,Guangdong,China)
出处
《中国生物医学工程学报》
CAS
CSCD
北大核心
2019年第4期392-400,共9页
Chinese Journal of Biomedical Engineering
基金
国家重点研发计划项目(2016YFC0104703)
