三维超声胎儿小脑分割与特征分析

发布时间：2019-03-21 17:32

【摘要】：超声成像技术具有无损、实时成像、价格低廉和使用便捷等优点,是产前诊断中最常用的检查手段。在常规的产前超声检查中,胎儿的小脑是必要的检查部位。小脑的横径和体积是医生评估胎儿发育状况和健康程度的重要指标,而小脑的形态和完整性是诊断先天性畸形的参考标准。然而,超声图像的低信噪比特性给定量分析和诊断带来了不利影响。临床应用中,超声诊断往往存在准确性因医生经验而异、三维特征分析耗时等缺陷。因此,在超声图像分析中引入自动特征提取和特征分析方法具有重要的意义。本论文的研究目标是开发一种创新的、高效的计算机辅助诊断方法,在无需人工干预的前提下,实现胎儿小脑三维超声容积的全自动分割和特征分析。基于此目的,本论文对超声图像的显著特征提取、目标结构定位、相位对称性提取及分割等方法进行了研究,工作内容包括以下四方面：1)实现图像显著特征的提取。在三维超声容积的中心切面上,胎儿的脑中线是最显著的特征,因为脑中线的灰度值较高、呈直线形状；其次是胎儿的头颅,面积较大、呈椭圆形。基于这些特征,本论文提出应用加权霍夫变换算法检测胎儿的脑中线,通过高斯函数改变图像的灰度分布,提高直线检测的准确性。此外,应用约束型随机霍夫变换算法提取胎儿的头颅，而通过约束点辅助椭圆拟合,提高算法的收敛速度。2)针对胎儿小脑面积较小、形状不规则以及边缘特征不明显等特征,提出了一种结合显著特征提取和遍历式搜索的间接定位方法。该定位方法将图像的显著特征和小脑的先验知识嵌入一种树形结构的概率型分类器,将小脑可能出现的位置约束在较小的范围内。然后,利用圆形滤波器进行遍历式搜索,对胎儿小脑进行精确定位。3)针对三维活动表面模型分割算法对初始模型敏感的缺点,将小脑的定位结果扩展至三维空间,作为分割的初始模型；针对该分割算法的弱边缘泄露问题,提出利用方向性相位对称性构造能量函数。方向性相位对称性将小脑的感兴趣区域分裂成若干个子区域,对各个子区域自适应地选择最匹配的滤波器角度,从而增强小脑边缘的清晰度和连续性、抑制图像的噪声,提高分割的准确性。4)基于分割的胎儿小脑三维模型进行特征分析,考察小脑左右半球的对称性以及小脑体积、横径和质心距与胎龄的相关性,并分别计算出相应的回归方程。
[Abstract]:Ultrasonic imaging technology has the advantages of lossless, real-time imaging, low cost and convenient use. It is the most commonly used means of prenatal diagnosis. In routine prenatal ultrasound examination, fetal cerebellum is a necessary part of examination. The transverse diameter and volume of the cerebellum are important indexes for doctors to evaluate the fetal development and health, and the shape and integrity of the cerebellum are the reference criteria for the diagnosis of congenital malformations. However, the low signal-to-noise ratio (SNR) characteristics of ultrasonic images have a negative impact on quantitative analysis and diagnosis. In clinical application, the accuracy of ultrasonic diagnosis varies with the doctor's experience, and the analysis of three-dimensional features is time-consuming and so on. Therefore, it is of great significance to introduce automatic feature extraction and feature analysis methods into ultrasonic image analysis. The aim of this paper is to develop an innovative and efficient computer-aided diagnosis method, which can realize automatic segmentation and feature analysis of three-dimensional ultrasound volume of fetal cerebellum without manual intervention. For this purpose, the methods of ultrasonic image salient feature extraction, target structure localization, phase symmetry extraction and segmentation are studied in this paper. The main contents are as follows: 1) realizing the extraction of image salient features. On the center section of three-dimensional ultrasound volume, the midline of the fetal brain is the most obvious feature, because the gray value of the midline of the brain is high, and the shape of the line is straight, and the next is the fetal head, with a larger area and an oval shape. Based on these features, this paper proposes to use weighted Hough transform algorithm to detect the midline of fetal brain, and to improve the accuracy of line detection by changing the gray distribution of the image through Gao Si function. In addition, the constrained random Hough transform algorithm is used to extract the fetal head, while the constraint point-assisted ellipse fitting is used to improve the convergence rate of the algorithm. 2) the fetal cerebellar area is smaller. Because the shape is irregular and the edge features are not obvious, an indirect localization method is proposed, which combines significant feature extraction and ergodic search. In this method, the salient features of the image and the prior knowledge of the cerebellum are embedded into a tree-structured probabilistic classifier, and the possible location of the cerebellum is limited to a smaller range. Then, the circular filter is used for ergodic search to accurately locate the fetal cerebellum. 3) aiming at the disadvantage that the 3D active surface model segmentation algorithm is sensitive to the initial model, the cerebellar localization results are extended to three-dimensional space. As the initial model of segmentation; In order to solve the weak edge leakage problem of the segmentation algorithm, an energy function is constructed by using directional phase symmetry. The directional phase symmetry divides the region of interest of the cerebellum into several sub-regions, and adaptively selects the best matching filter angle for each sub-region, thereby enhancing the clarity and continuity of the cerebellar edge and suppressing the noise of the image. 4) based on the segmented three-dimensional model of fetal cerebellum, the symmetry of the left and right hemispheres of cerebellum and the correlation of cerebellar volume, transverse diameter and centroid distance with gestational age were investigated. The corresponding regression equations are calculated respectively.
【学位授予单位】：复旦大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：R714.5;TP391.41

【参考文献】