脑MR图像颅骨分割算法

	脑MR图像颅骨分割算法
其他题名	The research of algorithm skull segmentation inMR images
	游佳丽
学位类型	硕士
导师	杨晓冬
	2015-10
学位授予单位	中国科学院大学
学位专业	光学工程
关键词	颅骨分割模板融合形变模型分类特征选择 C-v 模型
摘要	为了有效地从脑核磁共振图像中获取有价值的信息，研究者通常对图像进行自动处理与分析。而图像分割是所有图像分析中关键的一步。在过去的几年里，研究人员在脑提取和组织分割上，提出了许多相关的新算法，诸如BET， BEAST，LABEL 等。然而，在脑MR 图像上对脑颅骨进行的分割算法却很少有人研究。一方面，由于在MR 图像中颅骨的成像质量较差，另一方面，颅骨在MR 图像中呈现的状态不同，有时低密度，有时高密度，甚至与周围组织灰度相近。事实上，脑MR 图像的颅骨分割工作在某些领域中有很重要的应用。例如，脑电图和脑磁图的源成像问题和PET-MRI 一体机的衰减矫正工作。在构建源成像的正向矩阵问题中，因为人大脑的独特导电率分布，使用标准的人脑模型代替球模型，能提高其准确性。由于颅骨和软组织在导电率方面存在巨大差异，颅骨的分割在提高源成像的精度方面具有重要的作用。为了后续传导矩阵计算方便，要求颅骨部分呈现闭合状态。而目前的MR 图像颅骨分割算法不能处理闭合颅骨分割问题。本文提出的第一种颅骨分割算法能生成闭合的颅骨模型。基于弹性形变的算法的框架来源于由smith 等人在2002 年提出的BET 算法。本文的算法在BET 的原型基础上，变化曲面点的演变公式，分出颅骨内层和外层。我们将自动分割结果与手工分割结果进行对比，并与经典的方法进行比较，证明了本文方法的有效性。为了完成PET-MRI 一体机的衰减矫正工作，精确的分出MRI 图像中的颅骨部分起到了至关重要的作用。然而，一些相关的颅骨分割算法鲁棒性较差，比如2005 提出的基于形态学的分割算法。本文原创的提出基于模板融合和分类的颅骨分割算法，其鲁棒性较好，分割准确度较高。基于模板融合和分类的颅骨分割算法，思想来源于各种深度学习，分类的算法。由于分割的本身也是一个分类的过程，研究者将目光集中在基于分类的分割算法研究上。但由于颅骨在MR 图像中成像模糊，单纯使用分类的方法效果不理想。故在本篇论文中，又提出一种将模板融合作为先验信息，再进行分类的算法。通过与CT 图像中颅骨区域作为金标准的结果进行比较，该算法有较高的准确性。本文就颅骨的分割工作展开，第一章主要说明了课题的研究背景及意义。在第二章中，就传统的颅骨分割算法展开，主要介绍了基于边缘，区域和活动轮廓模型的分割算法。在第三和第五章中分别详细地介绍了提出的原创颅骨分割算法。在第四章中，介绍了离散粒子群优化算法在特征提取方面的内容，为第五章内容做铺垫。最后一个章节主要对全文进行总解。
其他摘要	In order to obtain valuable information from brain MR imaging, the first step of image processing is image segmentation. Researchers, who focused on the brain extraction and tissue segmentation in the past few years, proposed many related algorithms, such as BET, BEAST, LABEL, and so on, which increases the accuracy of brain extraction. However, not many algorithms on segmenting skull and scalp have been brought up due to few researchers concerning this problem. It is a challenge to propose an efficient skull segmentation algorithm. In one hand, the image quality of skull region in the MR image is poor. In other hand, it is difficult to distinguish skull from its neighborhood tissues because of similar intensity. In fact, skull segmentation in the MR imaging has been applied in various fields. The source imaging problem of EEG and MEG is a typical example. In the problem of constructing the forward matrix, using standard sphere model instead of individual brain model reduces the accuracy of source imaging because of the unique conductivity distribution of individual brain. To make the subsequent calculations simplified and feasible, it is required that the skull should be segmented as a closed part. However, recent existing skull segmentation algorithms of MR imaging cannot handle this request. So the first proposed skull segmentation algorithm focuses on generating a closed skull model to fit the subsequent relevant calculation steps. The frame of the deformable method is inspired from the brain extraction method proposed by Smith in 2002. Through changing the positions of the vertexes, the surface of the contour profile moves into interest area. Next, we compare the accuracy of segmentation of our method with the classic skull segmentation methods. The result shows that our method have a better performance in the problem of skull segmentation. To obtain the accurate attenuation correction of the PET-MRI all-in-one machine, it is vital to segment the exact skull part from the MR imaging. However, some related skull segmentation algorithms, like the one based on morphology proposed in 2005, are not robust to the images of MR. This article creates a new segmentation algorithm originated from the label fusion and the classification methods, which is of better robustness and higher segmentation accuracy. The algorithm of label fusion and classification is based on machine learning. Since the intensity of skull is uncertain in MR images, the method by using classification merely is hard to acquire the ideal results. In this paper, a new method, which set the label fusion as the prior information, was proposed. Our method receives a better performance in comparison with other two methods. First chapter had shown the background and the meaning of the research topic. In the second chapter, the classic skull segmentation algorithm was introduced. In the third chapter and fifth chapter, we proposed the new skull segmentation methods in detail. And in the fourth chapter, the principle of the feature selection method by using the discrete particle swarm optimization is explained. The last chapter has shown the conclusion of the paper.
语种	中文
文献类型	学位论文
条目标识符	http://ir.ciomp.ac.cn/handle/181722/49315
专题	中科院长春光机所知识产出
推荐引用方式 GB/T 7714	游佳丽. 脑MR图像颅骨分割算法[D]. 中国科学院大学,2015.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
游佳丽.pdf（2108KB）	学位论文		开放获取	CC BY-NC-SA	请求全文