胎儿腹部超声图像自动质量控制与参数测量

发布时间：2018-04-21 14:35

本文选题：胎儿超声 + 质量控制　；参考：《深圳大学》2017年硕士论文

【摘要】：产前超声检查过程中,超声图像的质量和相应生物参数测量的精度对于胎儿生长发育状况的准确诊断是非常重要的。但是,手动的质量控制是一个大量繁杂重复性工作的过程,临床上难以实现。另外,手动的解剖结构边缘轮廓勾画也非常耗时,在临床实践中会导致较大的医师间差异。为了提高产前检查的效率,减轻由不恰当的超声扫查切面选取和不精确的生物计量估计引起的测量误差,本论文提出全自动的胎儿超声图像质量控制(FUIQA)和参数测量(FUIPM)系统,协助实现临床产科检查中胎儿腹部超声图像的质量控制和精准的生物学参数测量。本论文提出的FUIQA系统由两个深度卷积神经网络模型实现,分别定义为L-CNN和C-CNN。L-CNN用于超声图像中胎儿腹部区域感兴趣目标(ROI)的定位。基于L-CNN定位到的ROI,C-CNN通过评价关键解剖结构胃泡(SB)和脐静脉(UV)描述的完整性来评估超声图像的质量。为了进一步提高L-CNN的性能,本文通过局部相位特征与原始超声数据通道融合的方式增加神经网络的输入源。通过全面的实验,阐明本文提出的FUIQA系统适用于16~40孕周的不同参数设置的胎儿超声图像,与超声医师手动评分相比达到90%的一致性。本论文提出的FUIPM系统是一个全自动分割胎儿超声图像的级联框架。首先,一个修正的全卷积网络(R-FCN)被用于提取多尺度视觉特征,产生密集的边缘预测图以鉴别解剖结构。为了提高预测图的局部空间一致性同时改善细节,本文将核心的R-FCN模型嵌入到自动上下文模型中,并将传统自动上下文模型中的并行级联改进为求和级联操作。大量的实验结果表明,FUIPM系统可以有效克服严重的边界不完整缺陷,相比其它流行的最新算法,可以取得更好的分割精度。本论文提出的FUIQA与FUIPM系统具有通用性,可以很容易的拓展到其他类型胎儿超声扫查切面的质量控制和参数测量,比如胎儿颜面部、四腔心切面等。
[Abstract]:During prenatal ultrasound examination, the quality of ultrasound image and the accuracy of biological parameters measurement are very important for the accurate diagnosis of fetal growth and development. However, manual quality control is a complex and repetitive process, which is difficult to achieve clinically. In addition, manual outline of the edge of the anatomical structure is also time-consuming, leading to greater differences between physicians in clinical practice. In order to improve the efficiency of prenatal examination and reduce the measurement error caused by improper selection of ultrasonic scanning section and inaccurate biometric estimation, a fully automatic FUIQA and parameter measurement system for fetal ultrasound image quality control (FUIQA) and parameter measurement (FUIPM) are proposed in this paper. Assist in the quality control of fetal abdominal ultrasound images and accurate measurement of biological parameters in clinical obstetrical examination. In this paper, the FUIQA system is implemented by two deep convolution neural network models, which are defined as L-CNN and C-CNN.L-CNN for localization of fetal abdominal area of interest (ROI). The quality of ultrasound images was evaluated by evaluating the integrality of the description of the key anatomical structures (SBB) and umbilical vein (UVS) based on the L-CNN localization of ROIC-CNN. In order to further improve the performance of L-CNN, the input source of neural network is added by the fusion of local phase feature and original ultrasonic data channel. Through a comprehensive experiment, it is clarified that the proposed FUIQA system is suitable for fetal ultrasound images with different parameters at 16 ~ 40 gestational weeks, which is 90% consistent with the ultrasound doctor's manual score. The FUIPM system proposed in this paper is a cascade framework for automatic segmentation of fetal ultrasound images. First, a modified full convolution network (R-FCN) is used to extract multi-scale visual features and generate dense edge prediction images to distinguish anatomical structures. In order to improve the local spatial consistency of the prediction graph and improve the detail, the core R-FCN model is embedded into the automatic context model, and the parallel concatenation in the traditional automatic context model is improved to sum concatenation operation. A large number of experimental results show that the FUIPM system can effectively overcome the serious boundary imperfections and achieve better segmentation accuracy than other popular algorithms. The FUIQA and FUIPM systems proposed in this paper are universal and can be easily extended to other types of fetal ultrasound scanning for the quality control and parameter measurement, such as fetal face, four-cavity tangent surface, and so on.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R714.5;TP391.41

【相似文献】