超声图像处理新方法及其在产前诊断中的应用

【作者】 余锦华；

【导师】 汪源源；

【作者基本信息】 复旦大学 ， 医学电子学， 2008， 博士

【摘要】 超声诊断具有无损、价廉、实时的优点,是现代临床医学必不可少的影像诊断技术之一。由于超声成像的相干特性,超声图像的信噪比低,给定量分析和诊断带来了不利影响。实际应用中,超声诊断往往存在准确性因医生经验而异、图像特征提取和分析方法滞后于成像技术发展等缺陷。在超声图像分析中引入自动特征提取和分析方法具有重要的意义,但要先解决图像降噪和图像分割两个关键问题。产科超声诊断是评价胎儿宫内生长发育的重要手段,对降低孕产妇和胎儿死亡率、确保优生优育具有重要意义。临床产科检查中,通过超声成像测量胎儿各径线长度从而预测胎儿体重是衡量胎儿宫内生长发育的重要指标之一。然而,目前临床胎重预测的准确性不尽人意,其主要原因有手动测量胎儿径线带来的随机误差和回归分析本身的系统误差。本文抓住降噪和分割这两个超声图像分析中的关键问题,以产科超声图像分析和辅助诊断为目标,在以下几个方面进行了研究和探讨:在图像降噪中,以各向异性扩散为主要研究对象。为高强度背景噪声中的噪声抑制和边缘检测提出核各向异性扩散模型;为增强斑点噪声抑制各向异性扩散滤波器SRAD的灵活性和稳定性,提出一种基于EM算法混合分布分解的扩散参数自适应选取方法;将非降采样Contourlet变换在几何结构捕捉方面的优势引入斑点噪声降噪中,提出一种称为Contourlet域复扩散的多尺度各向异性扩散模型。针对超声图像中造成图像信噪比低的斑点噪声和亮度不均匀,提出一种基于二维模糊聚类融合斑点噪声抑制和亮度补偿的超声图像分割方法,该方法可用于形状复杂的空腔和液体充盈组织的分割。将二维模糊聚类的思想进行扩展,提出针对脉冲噪声的快速二维模糊聚类算法和针对高斯噪声和亮度不均匀的多信息模糊聚类,分别用于自然图像和分子成像图的分割。针对超声图像中大量目标大体形状已知且变化不大的特点,提出一种融合边缘检测、模糊聚类、霍夫变换和形变模型的超声图像分割方法。该方法能够充分发挥每类算法的优势并避免它们的不足,可主要用于胎儿双顶径、头围、腹围和股骨长度的超声图像分割和参数测量。为抑制不准确测量数据对胎重预测模型训练的影响,将模糊逻辑引入支持向量回归中;为保证模糊支持向量回归的范化能力,引入遗传算法优化选取模型参数。将所提出的模型分别用于正常体重儿和低体重儿的胎重预测,与现有的回归方程法和人工神经网络模型进行对比,说明所提出模型的优越性。将胎儿参数自动分割与胎重预测模型结合起来,从参数获取和模型设计两方面避免临床胎重预测中存在的问题,建立一套完整的基于超声图像的胎重预测系统。该系统能提供较现有方法更准确的胎重预测,有潜力将预测误差降低40到70克。最后开发了一个产科超声图像辅助诊断系统。该系统具有数据库管理能力,并集成了论文中提出的所有图像降噪、图像分割和胎重预测方法。更多还原

【Abstract】 The ultrasonic diagnosis is one of the irreplaceable imaging diagnostic techniques in the modern clinical medicine, due to its merits of noninvasiveness, low cost and real-time imaging. Because of the characteristic of the coherent imaging process, ultrasound images suffer from a low signal-to-noise ratio that brings unfavorable effects into the quantitative analysis and diagnosis. In the pratical use, the ultrasound diagnosis confronts several shortcomings, such as the diagnosis accuracy strongly depending on the sonographer’s experience and methods of the image feature extraction and analysis falling behind the developement of imaging techniques. It is meaningful to introduce automatic feature exaction and analysis techniqes into the ultrasound image processing. However, two crucial problems, the image denoising and image segmentation should be firstly solved.As a major method to evaluate the fetal intrauterine growth, the obstetric ultrasound diagnosis plays an important role in decreasing the mortality of expectant mothers and fetuses, and thus in ensuring healthy pregnancy and scientific nurture. In the clinical obstetric examination, the estimated fetal weight （EFW） derived from ultrasonic measurements of fetal parts is one of important indices to estimate the fetal outcome. However, the accuracy of the clinical fetal weight estimation is unsatisfactory. Large random errors in manual measurements of fetal parts and system errors of the regression formula are two main causes of the inaccurate EFW.This dissertation focuses on these crucial problems in the ultrasound image analysis—image denoising and image segmentation, aims at the automated image analysis and computer aided diagnosis in the obstetric ultrasound. The studies have been carried out in following aspects.For the image denoising, we focus on the anisotropic diffusion technique. A kernel anisotropic diffusion model is proposed for the robust noise reduction and edge detection under the strong noisy background. To improve the flexibility and stability of the speckle reducing anistropic diffusion （SRAD） filter, a mixture distribution decomposion method is designed to estimate the diffusion parameters adaptively by using the EM algorithm. By making use of the good contour capturing ability of the nonsubsampled Contourlet transform, a multiscale anisotropic diffusion method called as the contourlet transform based complex diffusion is proposed for the ultrasonic speckle reducition.To alleviate the effects of speckle noise and intensity inhomogeneity in the ultrasound image segmentation, a novel method is proposed, which integrates the speckle noise reduction and the intensity inhomogeneity compensation into a two-dimensional homogenized fuzzy C-means （2DHFCM） framework. The 2DHFCM is applicable to the segmentation of cavities and fluid-filled tissues with a complex geometrical shape. By extending the principle of 2DHFCM, a fast two dimensional fuzzy C-means algorithm （Fast₂DFCM） is utilized into the segmentation of natural images corrupted by the implusive noise, and a multi-information based fuzzy C-means algorithm is utilized into the segmentation of molecular images corrupted by Gaussian noise and intensity inhomogeneity.According to characteristics that a large amount of objectives in ultrasound images appear similar geometrical shapes and limited deformation, a method is developed by integrating the edge detection, fuzzy clustering, Hough transform and active contour model. The proposed method is able to make the best use of the strength of different segmentation algorithms, while avoiding their deficiencies. This method can be used in the segmentation and measurement of fetal biparietal diameter, head circumference, abdominal circumference and femur length.The fuzzy logic is introduced into the support vector regression to limit the contribution of suspect inaccurate measurements to the training of the fetal weight estimation model. To guarantee the generalization performance of the fuzzy support vector regression （FSVR）, the genetic algorithm is employed to obtain optimal paramters for the FSVR. The proposed FSVR modle is used in the weight estimation respectively for normal weight fetuses and low birth weight fetuses. Its performance has been demonstrated by the comparison with the existed regression formulas and artificial neural network models.To avoid shortcomings associated with parameters acquisition and the model design in the clinical fetal weight estimation, a complete fetal weight estimation system is developed by integrating the automated fetal ultrasound image segmentation with the newly developed models. The fetal weight estimation system can provide more accurate fetal weight estimation than existed methods and decrease the estimation errors around 40 to 70 g. Finally, a computer aided diagnosis system is developed for obstetric ultrasound images. The system contains functions of the database manipulation and integrates the image denoising, image segmentation and fetal weight estimation methods which have been previously developed.更多还原