【作者】 张胜君；

【导师】 陈后金；

【作者基本信息】 北京交通大学 ， 电路与系统， 2013， 博士

【摘要】 摘要：乳腺癌的早期检测和诊断是挽救患者生命的最有效途径。目前乳腺X线图像是乳腺癌检查最主要的影像方法,但是早期具有隐匿性的乳腺癌影像特征一般不够明显,受医生主观方面的影响,极易出现误诊和漏诊的情况。随着计算机技术的不断发展,乳腺癌计算机辅助检测已经成为医学图像处理领域的一项研究热点,有效的计算机辅助检测方法可以辅助医生更好的分析乳腺X线图像,从而提高乳腺癌诊断的准确率。微钙化点、肿块、结构扭曲是乳腺癌在X线图像上的主要表现,对这些病灶的检测是计算机辅助检测领域的一项研究难点。本文在前人研究基础上,通过结合机器学习、计算机视觉等领域的思想,构建乳腺癌的计算机辅助检测系统。本文取得的创新性研究成果总结如下：(1)深入研究了乳腺X线图像乳腺区域分割方法和预处理中图像去噪,实现了基于最佳阈值的乳腺区域分割方法和基于自适应中值滤波的乳腺X线图像平滑方法。最佳阈值分割方法通过计算最佳阈值,快速有效的剔除乳腺X线图像中的标签、背景,分割出乳腺区域,从而确定了后续处理过程的研究区域；自适应中值滤波方法在进行滤波处理时为滤波器提供负反馈,在滤除噪声的同时能较好的保留图像细节信息,为乳腺X线图像后续处理提供了前提条件。(2)针对乳腺X线图像微钙化点检测中微钙化点分割不够完整的问题,提出一种基于多分辨率区域生长和图像差值的微钙化点分割方法。通过改进传统区域生长的生长规则,在每一个目标区域得到一个最佳阈值,使每个目标区域均得到较好的分割效果,有效的保持了微钙化点的完整形状和分布信息,为后续的微钙化点检测提供了支持。(3)深入分析了相关向量机及其最新进展——自适应核学习相关向量机的理论及推导过程,利用自适应核学习相关向量机的核参数自动优化设置和模型更稀疏的特性,探索性地将其应用于乳腺X线图像中肿块的检测,提出了一种基于自适应核学习相关向量机的乳腺x线图像肿块检测方法。实验结果表明该方法较支持向量机、相关向量机方法具有更好的检测性能和鲁棒性。(4)针对乳腺X线图像结构扭曲检测假阳性率偏高的问题,提出了一种新的乳腺x线图像结构扭曲检测方法——相似度收敛指数(Similarity Convergence Index,SCI)方法。首先利用马氏距离比计算出毛刺的相似度,然后通过计算相似度加权的收敛指数增强放射状毛刺,最后提取出收敛指数的局部最大值作为候选点,并对这些候选点进行分类,检测出结构扭曲。实验结果表明该方法可有效的减弱非毛刺结构对检测的影响,大幅度降低结构扭曲检测的假阳性率,同时对于不同类型的乳腺X线图像具有较强的鲁棒性。更多还原

【Abstract】 ABSTRACT:Early detection and diagnosis of breast cancer is the most effective way to save lives of patients. Mammogram is the most preferred method for breast cancer census at Present. However, the ambiguous characteristics of the early cancers and the subjective impact of the doctors will all probably induce the error and miss on diagnosing. With the rapid develop of computers, computer-aided detection of breast cancer has become one of the hot research points among medical image processing field. The effective computer-aided detection method can help doctors realize and analyze the mammograms better, and further improve the accuracy of diagnosis. Micro-calcifications(MCs), mass and architectural distortion(AD) are major signs of breast cancer, while the detection of these lesions is a difficult and challenge problem of the computer-aided detection system. The paper is established based on the previous research, and through incorporating the ideas of machine learning and computer vision, to construct the computer-aided detection analysis system. The main contributions and innovation are as follows:(1) Breast region segmentation and image denoising in mammogram preprocessing are explored. A breast region segmentation method based on optimal threshold and an image smoothing method based on adaptive median filtering are realized. The optimal threshold is computed to remove labels and background from mammogram to segment breast region, which can obtain research region for subsequent processing. Adaptive median filtering provides negative feedback for filters, which is capable of smoothing noise as well as retaining the details in the image, providing support to subsequent processing.(2) In view of the problem that micro-calcifications are segmented incompletely in micro-calcification detection, a method of micro-calcification segmentation based on multi-resolution region growth and mage difference is proposed. In the method, an optimal threshold is obtained in each target region, in which the segmentation is effective. The shape and distribution of micro-calcifications can be obtained accurately, providing support to the subsequent micro-calcification processing.(3) The basic theories of relevance vector machine (RVM) and its development---adaptive kernel learning based relevance vector machine (aRVM) are analyzed. Using the capable of aRVM automatically learning the parameters of the kernel, the use of aRVM for detection of mass in mammograms is explored, and a detection method for mass based on aRVM is proposed. The experimental results show that for mass detection, the aRVM method has better detection performance and robustness than method of SVM and RVM.(4) For the problem of high false positive rate in the detection of architectural distortion (AD) in mammograms, a method to detect AD based on speculation similarity convergence index (SCI) is proposed. In the method, the spiculation similarity based on Mahalanobis distance is presented and applied to compute SCI to enhance radiating spiculations. Then local maximum values of SCI are extracted as AD candidates, and lastly these candidates are classified into AD and normal tissues. The experimental results show that the proposed method can effectively weaken the effect of non-spiculations on AD detection, which can reduce false positive rate significantly and be applied to mammograms of various breast types.更多还原