【作者】 高梁；

【导师】 陈武凡；

【作者基本信息】 电子科技大学 ， 检测技术与自动化装置， 2013， 博士

【摘要】 乳腺癌是目前女性疾病中最常见的恶性肿瘤，已成为导致女性死亡的主要杀手之一。早发现、早诊断和早治疗是目前医学上对防治乳腺癌采取的“三早”原则。超声成像凭借其无创伤、无辐射和费用低廉等优点，已成为乳腺肿瘤临床诊断的主要手段之一。乳腺超声图像的肿瘤分割可以给医生提供辅助诊断和参考意见，提高了诊断的客观性和正确性，降低了误诊和漏诊现象的发生。然而，乳腺超声图像的准确分割非常困难。首先，在超声成像过程中，由于受成像设备的影响，超声图像产生了固有的斑点噪声，使得图像的信噪比和对比度较低、边缘模糊甚至边缘信息缺失。其次，超声图像中存在多种伪影，比如衰减以及在射频场中，非均匀光束在人体内衰减导致的伪影。最后，不同乳腺肿瘤的大小、形状和位置差异较大。除此之外，乳腺超声图像中存在肿瘤的浸润效应，也就是，肿瘤经常侵入周围的正常组织，这使得肿瘤与周围正常组织，比如皮下脂肪组织，腺体组织非常相似，往往难以分辨，这给肿瘤分割造成了更大的困难。乳腺超声图像的上述特点使得肿瘤的分割仍然非常具有挑战性，是一个值得深入研究的课题。为此，本文在图论、曲线演化理论和水平集方法的基础上，深入、细致地研究了基于图的Normalized Cut方法、水平集活动轮廓方法以及它们在二维或三维乳腺超声图像分割中的应用。本文的主要工作和研究成果如下：1.基于相位和梯度矢量流的水平集乳腺超声肿瘤分割在距离正则化水平集模型的基础上，本文提出一种基于相位和梯度矢量流的水平集活动轮廓模型。首先，引入单演信号的概念，利用柯西核而非Log Gabor构造正交滤波器，以提取多尺度的基于单演信号的特征；然后，将相位非对称方法应用于多尺度的图像特征上，以达到检测边缘和去除噪声的目的；接着，利用得到的边缘图，定义了一个边缘停止项，并在此基础上改进了梯度矢量流；最后，在活动轮廓模型中融入自定义的边缘停止项和梯度矢量流。一方面，由于本文方法仅利用了局部相位信息，因此它对乳腺超声图像中存在的灰度不均匀性和噪声具有较好的鲁棒性。另一方面，梯度矢量流的引入使得本文方法能够有效地捕获凹陷边界和弱边界。乳腺超声图像的分割实验证明，本文方法可以成功地分割出乳腺肿瘤。2.基于同质片的乳腺超声图像肿瘤分割在Normalized Cut（NCut）方法的基础上，本文提出了基于同质片的乳腺超声图像肿瘤分割方法。首先，定义了基于灰度和纹理信息的边界检测函数；然后，利用边界图定义了一种可以随图像像素位置变化而变化的自适应邻域系统，称之为同质片。由于同质片可以保证邻域不会跨越不同的组织，因此基于同质片的统计特征更有利于区分不同的组织，进而能够提高分割的准确度。最后，将每个同质片视为一个模糊集合，用同质片内纹理基元的模糊分布作为分割特征，再用NCut方法得到乳腺肿瘤。本文提出的算法既可以避免超声图像衰减伪影的影响，同时也避免了将肿瘤周围的正常组织，比如皮下脂肪组织、腺体组织等误划分为肿瘤。100例乳腺超声图像上的分割实验证明，本文算法取得了比现有方法更高的分割精度和鲁棒性。3.基于同质体和局部能量的三维乳腺超声肿瘤分割在经典的水平集活动轮廓模型Chan-Vese模型的基础上，借鉴Lankton方法局部处理的思想，本文提出了一种基于同质体和局部能量的水平集活动轮廓模型，并用于分割三维超声图像的乳腺肿瘤。首先，用三维相位非对称方法提取边缘曲面，进而利用每个体素邻域的边缘信息拟合了一个二次曲面，将拟合后曲面的极值作为各体素的边缘能量。这样，既使得提取的边缘曲面更为连续和光滑，而且也去除了曲面表面的噪声。然后，将二维同质片的概念推广至三维空间上，利用拟合后的边缘曲面构造了可以保持局部同质性的三维自适应邻域，将其称之为同质体。接着，通过分析同质体的纹理分布和计算局部灰度均值，定义了演化曲线周围基于同质体的局部能量。和同质片类似，利用基于同质体的特征有利于区分局部表现和外观相似却属于不同组织的体素。同时，还定义了演化曲线周围基于全局灰度均值的全局能量，这有利于驱动离真实肿瘤边界较远的演化曲线逐渐向真实边界处靠拢。最后，在活动轮廓模型中，综合全局和局部能量，通过曲线的快速演化实现乳腺超声图像的肿瘤分割。所提出的方法既可以克服乳腺超声图像灰度不均的问题，也可以避免陷入局部极值的问题。25例三维乳腺超声图像上的分割实验证明，与现有的方法相比较，所提出的方法可以有效地分割出乳腺超声肿瘤。更多还原

【Abstract】 Breast cancer is the most common malignant lesions and is the leading cause ofdeath in women. Prevention and treatment of breast cancer adopts three principlesclinically: early detection, early diagnosis and early treatment. Ultrasound (US) imaginghas been one of the main measures of clinical diagnosis on breast tumor due to itsnon-invasive nature, minimal ionizing radiation and low cost. Tumor segmentation ofbreast US image can provide aided diagnosis and second opinion. This improves theobjectivity and accuracy of diagnosis and reduces the likelihood of misdiagnosis andmissed diagnosis.However, accurately segmenting breast tumors in US images is a very difficult taskdue to the following reasons. First, US image has inherent speckle, fuzzy or evenmissing edges, low signal-to-noise ratio, and low contrast due to the effect of USimaging device. Second, there are characteristic artifacts, such as attenuation and thosecaused by nonuniform beam attenuation within the body in radio-frequency field. Third,tumor variance in shape, size and location differs greatly. The effect of tumorinfiltrating is also a reason, that is, the tumor often infiltrates into its surroundingnormal tissue. This leads to the presence of tumor-like structures in malignant tumorimage such as, subcutaneous fat and glandular tissue. It is difficult to distinguish themalignant tumor from these tumor-like structures visually and hence, tumorsegmentation task is much more difficult. All characteristics of breast US images givenabove make US image segmentation very challenging. Segmentation of ultrasonictumor is worthy of further study.Based on the graph theory, curve evolution theory and level set method, this papermakes in-depth study on Normalized Cut (NCut) graph-based method, active contourlevel set method and their applications in two-dimensional(2-D) orthree-dimensional(3-D) US image segmentation. The main contents and innovationinclude the following three aspects:1. Phase-and GVF-based level set segmentation of ultrasonic breast tumorsBased on the distance regularized level set evolution model, this paper presents a phase-and GVF-based level set method for segmentation of ultrasonic breast tumors.First, by introducing the concept of monogenic signal, we use Cauchy kernels ratherthan Log Gabor as pair of quadrature filters for the multi-scale feature extraction.Second, phase asymmetry approach is then applied to multi-scale features to enhanceedges and remove noise effect. Third, based on precalculated edge map, an edgestopping term is defined and gradient vector flow (GVF) is then improved. At last, theedge stopping term and the resulting GVF are incorporated into the active contourmodel. The proposed method is insensitive to intensity inhomogeneities and noise dueto the use of local phase information. On the other hand, the proposed method cancapture concave boundaries and weak boundaries due to the use of GVF field.Experiments on clinical breast US images showed that the proposed method can extracttumor boundaries from breast US image, as compared to the state-of-the-art methods.2. Segmentation of ultrasonic breast tumors based on homogeneous patchThis paper presents a novel algorithm based on homogeneous patches (HP) andNCut for segmentation of breast tumor in US images. A novel edge-detection functionis defined by combining intensity and texture information to look for boundaries in USimages. Subsequently, a kind of adaptive neighborhood according to image locationreferred to as homogeneous patch (HP), is proposed by using edge map from anedge-detection function. The HPs are guaranteed to spread within the same tissue region.Hence, the statistic features within HPs can better distinguish different tissues and,furthermore, improve accuracy in image segmentation. Each HP is considered as afuzzy set. The fuzzy distribution of textons in HPs is used as final image features andtumor segmentation is obtained by using the NCut method. The proposed method canavoid attenuation artifacts and decrease the likelihood that the surrounding structuresare misclassified as tumor. Experimental results from100breast sonogramsdemonstrated the improvement in accuracy and robustness in segmenting the breastultrasound images by the presented algorithm, as compared to the state-of-the-artmethods.3.3-D segmentation of ultrasonic breast tumors based on homogeneous volumeand local energyBy taking the advantages of the classical Chan-Vese (CV) level set model andLankton method, this paper presents a novel level set active contour model based on homogeneous volume and local energy for3-D segmentation of ultrasonic breast tumors.First,3-D phase asymmetry approach is used to extract edge surface. And then at eachvoxel in3-D image, a quadric surface is fitted to estimate the edge energies based on theprecalculated edge surface. The fitted surface is continuous and smooth and isinsensitivity to noise. Second, based on the precalculated fitted surface, the concept of2-D homogeneous patch (HP) is extended to3-D homogeneous volume (HV), i.e., akind of adaptive neighborhood which can guarantee locally homogeneous neighborhoodin3-D. Third, local energies are defined at each voxel along the curve by usingdistribution of textons and mean intensity within HVs. Similar to HPs, using HVs helpsto discriminate those pixels with similar appearance but belonging to different tissues.At the same time, global energies are defined at each voxel along the curve by usingglobal statistics. When the contour is far from object boundaries, the force from theglobal energies is used to guide the contour toward and finally stops the contour atobject boundaries. At last, the local and global energies are incorporated into thegeometric active contour model and tumor segmentation is obtained by using the fastcurve evolution method. The proposed method can overcome the problem of intensityinhomogeneities and avoid falling into local extrema. Experiments on25clinical3-Dbreast US images showed that the proposed method can segment3-D ultrasonic breasttumors accurately, as compared to the state-of-the-art methods.更多还原