基于钼靶图像的计算机辅助乳腺癌检测系统中关键技术研究

【作者】 刘俊；

【导师】 陈建勋； 唐金山；

【作者基本信息】 武汉科技大学 ， 机械电子工程， 2012， 博士

【摘要】 乳腺癌是女性最常见的恶性肿瘤之一，早期的检测能极大地降低死亡率。钼靶成像技术采用低剂量X射线检查妇女的乳房，是临床上女性乳腺癌疾病最常用的检测手段。医生通过观察钼靶图像中的钙化、肿块等异常区域来诊断妇女乳腺癌，但是图像的阅读对医生的临床经验要求较高，诊断结果往往也会受主观因素的影响，因此研制可靠的计算机辅助诊断系统具有重要的现实意义。本文主要对基于钼靶成像技术的计算机辅助癌症诊断系统的一些关键技术进行了研究，主要开展的工作如下：1）图像增强是计算机辅助癌症诊断系统中得到医生肯定的一种技术并得到了广泛的应用。本文对JPEG压缩图像的增强技术进行了研究，提出了一个新的基于DCT域的JPEG压缩的图像的增强算法。在该算法中，根据用户给定目标的对比度值和视觉质量要求，先增强每个DCT块，再将整个图像进行解压，采用遗传算法搜索最优参数设置来对图像进行增强。通过客观测试和主观测试，这种新的算法有效减少增强效果所带来的边块效应，大大提高了医生对乳腺非正常区域的辨别。2）目前肿块的分割多采用人工方式或半自动分割方式，人工分割方式效率不高，半自动分割方式也需要人工干预，本文将两种常用的图像分割方法进行了结合，提出了一种全自动的乳腺肿块分割算法。该算法先用标记分水岭算法对乳腺肿块进行粗分割，然后使用本文改进的水平集活动轮廓方法对肿块进行精确分割。分水岭分割运行速度快，水平集方法分割精准，新的算法结合这两种方法的优点，加快了整个分割处理的过程，并提高了分割效率。此外，该方法具有良好的拓扑适应性，它可以处理形状较为复杂的乳腺肿块。3）典型的良性肿块具有圆形、平滑且清晰的边界特征，而恶性肿瘤通常具有多刺、粗糙且模糊的边界特征，边界的特征是肿块良恶性诊断的重要依据。乳腺肿块区域被分割出来以后，除了提取肿块的统计特征、几何特征和纹理特征等常用特征以外，本文提出了梯度信息中的一套新特征。该特征由乳腺肿块边界及肿块与图像背景间带状区域所提取，用来表达基于轮廓像素相对梯度走向的突刺结构，此类特征的增加提高了分类结果的准确率。此外，本文在经典LBP（局部二进模式）的基础上提出了ILBP（改进的局部二值模式）算子。该算子将图像块的中位数作为新的阈值，并且保持了中心像素值的信息。从1×1到9×9的图像块提取ILBP特征以后的分类结果表明使用新特征的分类准确率比使用原始的LBP特征的分类准确率提高了5%左右。4）肿块分类可作为计算机辅助诊断乳腺癌的重要依据。目前乳腺肿块的分类主要基于单分类器算法或改进算法。为了满足对肿块特征普适性和鲁棒性的要求，本文将诸多特征融合来训练分类器；同时研究了各种不同的分类算法来识别肿块的良恶性，包括LDA（线性判别分析）+KNN（K最邻近结点算法），RF（随机森林）算法和SVM（支持向量机）等;并在大型数据样本集中对上述算法进行了评价和测试，为形成高精度的乳腺癌肿块异常区域检测和良恶性识别算法提供坚实的基础。5）数字化技术的显著进步和医学影像数量的几何级增长，医生对相似病例的图像查阅工作变得更加困难和耗时。本文还设计并实现了一个基于纹理特征的乳腺癌肿块图像检索系统；该系统主要根据输入图像的纹理特征，计算输入图像和样本图像的相似度，实现了特征提取、查询、匹配、显示等功能。系统仿照医生诊断的一般流程，返回的相似参考病例图像，有效地辅助医生从大量的经验数据中得到经验值，从而对当前病例的确定提供帮助。更多还原

【Abstract】 Breast cancer is the most common malignancy in females. Early diagnosis and treatmentcan increase the survival rate. Mammography is currently the most effective method to detectearly breast disease, by using the low dose X-ray to check the breasts. The radiologists diagnosebreast cancer according to the abornalmality in mammography such as micro-calcifications andmasses. For the radiologistes, however, great clinical experiences are requested to read themammography images, as well as their diagnosis could be affected by many subjective factors.Therefore, it is necessary to develop reliable computer aided diagnosis (CAD) to overcome suchlimitations.In this thesis, we investigated the key CAD techniques related to calcification and massdetection and diagnosis in mamography and solved some problems from current algorithm inbreast cancer detection. The related works are carried out as follows:1. Image enhancement is widely used in calcification detection. But current algorithm couldaffect the correct diagnosis by enhancing imate features and noises together. In order to processthe mammography image compressed by JPEG, we investigated the image enhancementtechnique based on DCT domain and proposed a new algorithm. In this algorithm, each DCTdomain is enhanced firstly according to the contrast value and visual quality given by the user.Then the whole image is decompressed and enhanced by the optimized parameters from geneticalgorithm. This new algorithm can decrease the artifact due to enhacement effection, which hasbeen tested by objective and subjective detection, which can significantly improve theidentification of calcification in breast by radiologists.2. Manual and semi-automatic segmentation are mainly used in mass segmentation.However, manual segmentation lack efficiency, semi-automatic method still needs the manualintervention. We tried to combine the two common methods together and proposed a fullyautomatic algorithm for breast mass segmentation. The mass is roughly segmented by themarked watershed algorithm, and then precisely segmented by our improved level set activecontour. The new algorithm combines the fast running speed of watershed and precisesegmentation of level set, which ensure the fast and efficient segmentation. Moreover, thesealgorithms possess good topological flexibility, enabling segmentation of mass with complicatedcontour.3. It is known that a typical benign mass has a round, smooth, and with well-circumscribedboundary, while a malignant tumor usually has a spiculated, rough and blurry boundary. Thus, boundary analysis has been widely used for the benign or malignant classification of masses.After the segmentation of mass region, statistics, geometry and texture features have beenextracted. Other than these, we proposed a group of new features from gradient information,which are extracted from he boundary of the mass and margin region between mass andbackground to express the spiculation based on the relative gradient orientation of mass contourpixels. Such features could increase the accuracy rate. We also analyzed the texture features andproposed an improved local binary pattern (ILBP) operater on the basis of classical local binarypattern (LBP). ILBP operater chooses the median of image block as the new threshold andmaintains the information of center pixel value. The classification resutls based on the ILBPfeatures extracted from1×1to9×9image blocks showes that, the accurate ration forclassification based on new features increases5%comparing to that on the original LBPfeatures.4) Mass classification is the major support for the CAD in breast cancer diagnosis. Mostmass classification is made on the single or improved classifier, a few is on the integratedclassifier. To satisfy the requst of universality and robustness for mass features, we trained theclassifier by integrated features. The single classifier by machine learning methods isinverstigated, including linear discriminant analysis (LDA) and support vector machine (SVM).We also tested and estimated the above methods in the large data set. The random forest (RF) isfirst used for mass classification. Our work provides solid support for the precise detection andclassification of abnormality in breast mass.5) On the basis of the above research, we design and realize a breast mass images retrevialsystem based on texture features. This system cans calcualte the similarity between imput imageand sample image according to the texure features of input image. Features extraction, inquiry,matching and display can function well in this system. It provids the visual assistance toradiologists by mimicing their diagnosis process and return similar reference case image.更多还原