妇产科早期诊断中超声图像的分析与处理研究

【作者】 邓寅晖；

【导师】 汪源源； Yuzhong Shen；

【作者基本信息】 复旦大学 ， 医学电子学， 2011， 博士

【摘要】 超声诊断是目前最重要的医学影像诊断方式之一,具有无损、实时、非侵入式等特点,在妇产科诊断中得到了广泛应用。妇产科超声诊断能有效检测母体和胎儿的生理健康,是提高诊断准确率与客观性的重要手段。妇产科超声诊断现阶段的一个研究热点是针对疾病早期的检测,它能有利于进一步的医学处理,但同时又增大了诊断的难度。目前临床超声诊断主要存在对医生经验依赖性高、检测过多人工操作以及信息处理相对困难等问题,这对临床诊断中超声图像自动分析、处理的研究提出了要求。由于超声成像过程中回波间存在相互干扰,形成的斑点噪声影响了进一步图像分析、处理算法的性能,因此斑点噪声的处理以及目标模型的建立是超声图像分析、处理研究必须要考虑的两个基本问题。本论文针对妇产科早期诊断中的三个热点问题,以计算机辅助诊断为目标,对超声图像的分析、处理方法进行研究。针对多囊卵巢内小囊胞的检测问题,提出自适应形态滤波的斑点噪声去除方法,该方法能对斑点噪声在图像上表现出的像素值突变进行有效抑制：提出改进的带标记分水岭算法,分割出候选的小囊胞,并提出基于剩余谱估计的感兴趣区域自动选取算法；进一步提出目标生长算法,实现了对卵巢内小囊胞的自动识别。实验表明该方法的识别率和误识率分别为89.4%和7.45%,与其他方法的比较验证了该方法的优越性。针对胎儿早期心脏结构的检测问题,提出基于运动累加图的感兴趣区域自动选取方法,该方法能在超声图像序列信噪比较低的条件下对胎儿心脏区域进行准确提取；针对早期胎心图像的低信噪比以及检测时对运动信息的需要,提出瑞利均值各向异性扩散方法；为准确检测早期胎儿心脏结构,进一步提出能同时考虑胎儿心脏结构与运动信息的活动心脏模型,通过优化模型的结构保证了推断过程的收敛性。在胎儿早期心脏结构检测的实验中,90%的检测结果其误差小于13个像素。针对颈项透明层的自动识别问题,提出一个具有三层结构的分层模型。该模型针对颈项透明层检测的具体需要,采用基于梯度方向直方图特征的支撑向量机来表征图像上的相关目标,并建立相应的空间模型来描述不同目标间的相互关联,通过动态规划以及广义距离变换对模型的全局最优解进行求解。实验结果表明,该模型对颈项透明层的自动识别可以达到约60%的准确率。更多还原

【Abstract】 Ultrasound diagnosis is one of the most important diagnostic imaging techniques, which has been widely applied in the diagnosis of gynecology and obstetrics due to its merits of harmlessness, real-time imaging and noninvasiveness. Ultrasound diagnosis of gynecology and obstetrics can be effectively used to evaluate the physiologic health of expectant mothers and fetuses, which is the important way to improve the accuracy and objectivity of the diagnosis. Nowadays one hot research topic of ultrasound diagnosis of gynecology and obstetrics is the detection for the early disease. This early diagnosis can benefit the further medical process but also increases the difficulty of diagnosis. The current problems of such clinical ultrasound diagnosis mainly come from the high dependence of the experience of doctors, the manual operations for detections and the difficulty of processing the information. Therefore it is required for the researches to set up an automatic analysis and processing system for ultrasound images.Because of the interferences of ultrasound echo signals during the imaging, the formed speckle may affect the performance of further image analysis and processing algorithms. Therefore the two key problems of the analysis and processing methods for ultrasound images are the processing of speckle and the establishment of object models.This dissertation focuses on three hot topics of the early diagnosis of gynecology and obstetrics. To automatically aid the diagnosis, the studies are carried out on the analysis and processing methods for ultrasound images.For the detection of follicles of polycystic ovary syndrome, the adaptive morphological filtering is first proposed for despeckling, which can effectively depresse the abrupt variations of pixel values due to the speckle. Then the enhanced labeled watershed algorithm is proposed to segment the candidates of follicles and the iterative algorithm of automatically selecting the region of interest is proposed based on the spectral residual approach. Finally the object growing algoritm is proposed to realize the automated detection of follicles inside the ovaries. Compared with other methods, experimental results demonstrate the better performance of this proposed method which achieves the 89.4% recognition rate and 7.45% misidentification rate. For the detection of the structure of early fetal hearts, the automatic selection algorithm for the region of interest is first proposed based on the accumulated motion image. This algorithm can overcome the low signal-to-noise ratio of the ultrasound image sequences and accurately extract the region of fetal hearts. Then the Rayleigh-trimmed anisotropic diffusion is proposed to deal with the low signal-to-noise ratio of the ultrasound images of fetal hearts and emphasize the motion information for the next processing. Finally the active cardiac model is proposed for the automated detection of early fetal cardiac structure. Both the structure and the motion information of fetal hearts are considered simultaneously with this model. The convergence of the inference of the model is guaranteed through optimizing the structure of the model. According to the experimental results of detecting the early fetal cardiac structures, the error of 90 percent of detections of this proposed method is less than 13 pixels.For the automated detection of the nuchal translucency region, a three-layer hierarchical model is proposed. In this model, the support vector machine is applied to represent the related objects based on the features of the histogram of oriented gradient. Then the corresponding spatial model is established to define the spatial constrains between different objects. The optimal solution for the inference of the model can be obtained by applying the dynamic programming and generalized distance transform. Experimental results demonstrate that the model can achieve the accuracy of about 60% for the automated detection of the nuchal translucency.更多还原