基于多模态MRI的脑白质病变检测算法研究

发布时间：2018-06-28 07:04

本文选题：脑MRI + 多模态　；参考：《杭州电子科技大学》2014年硕士论文

【摘要】：脑白质疏松症是一种在T2FLAIR的MRI等医学影像中双侧脑室前后角周围白质、半卵圆中心、放射冠等区域所表现出的大小不等的斑点状或融合成片状的白质高信号区，是中老年的常见病之一。它的病理学成因到现在为止还并不是很清楚，可能与中风、帕金森病、轻度认知障碍、阿尔茨海默氏症、甚至情绪紊乱等有联系。同时，在与患者T2FLAIR影像中的白质疏松区域对应的ESWAN影像的相应区域中可以观察到脑内深部髓质静脉的显现，它们之间的关系近年来开始被医学界探索，研究者正试图通过它们在影像学上的关系和表现进一步解释其关联性。临床医生往往通过主观经验或者简单的体视学网格测量对脑白质疏松症程度进行评级；而脑内深部髓质静脉即使在最明显的ESWAN影像中仍是管径细小、成像模糊、分辨困难。这些主观评价无法提供一个准确的论据，因此需要一种可靠的计算机辅助方法帮助医生准确提取脑白质疏松症病变区域和脑内深部髓质静脉。针对这样的需求，，本文根据两种病变区域在多模态影像下的成像特点，研究了对T2FLAIR影像中脑白质疏松症和ESWAN影像中脑内深部髓质静脉主干两者进行自动提取的计算机辅助诊断方法，主要分为三大步：先对T2FLAIR影像进行头骨剥离预处理；然后在T2FLAIR影像中提取脑白质疏松区域；最后在ESWAN影像的相应区域（与T2FLAIR影像中的脑白质疏松区域对应）提取脑内深部髓质静脉主干。头骨剥离预处理先后利用了大津阈值法和形态学方法，可以对干扰脑白质疏松分割的非大脑组织高信号进行有效剥离，实验中本方法与通用的BSE头骨剥离方法一起与手动精细分割金标准相比，在Jaccard相似度和Dice相似度两个测度上得出本方法上更为准确。脑白质疏松区域分割则首先通过大津阈值法确定脑组织中脑白质疏松区域的初始轮廓，然后用一种改进的C-V模型方法对初始轮廓进行迭代演化，对比实验证明本方法比Li的免初始化的水平集方法在分割效果上表现略好，在分割时间上表现更优。最后，利用了Hessian矩阵的特征值与脑内深部髓质静脉灰度变化方向之间的关系设计了一种筛选器，用于脑内深部髓质静脉主干的提取，该筛选器能够初步筛选出有效主干，再结合形态学后处理连接断开主干，剔除伪主干，结果与手动精细分割的金标准比较接近。
[Abstract]:Leukoaraiosis is an area of white matter in the white matter around the anterior and posterior angles of the ventricle, the center of the semiovale, the corona, and other regions in the MRI images of T2FLAIR. It is one of the common diseases in the middle and old age. Its pathogenetic causes are still unclear and may be associated with stroke, Parkinson's disease, mild cognitive impairment, Alzheimer's disease, and even emotional disorders. At the same time, the deep medullary veins in the brain can be observed in the corresponding regions of ESWAN images corresponding to the areas of leukoaraiosis in T2FLAIR images of patients. The relationship between them has been explored by the medical community in recent years. The researchers are trying to further explain their relevance through their imaging relationships and performance. Clinicians often rate the degree of leukoaraiosis by subjective experience or simple stereological grid measurements, while deep medullary veins in the brain are still small in diameter, blurred and difficult to distinguish in the most obvious ESWAN images. These subjective evaluations can not provide an accurate argument, so a reliable computer-aided method is needed to help doctors accurately extract the diseased areas of leukoaraiosis and the deep medullary veins in the brain. In view of this demand, according to the imaging characteristics of two kinds of pathological regions in multi-mode images, The computer-aided diagnosis method for automatic extraction of the main trunk of deep medullary vein in T2FLAIR image and ESWAN image was studied. It was divided into three major steps: first, the skull stripping preprocessing of T2FLAIR image was carried out; Then the loose area of white matter was extracted from T2FLAIR image and the main trunk of deep medullary vein was extracted from the corresponding region of ESWAN image (corresponding to the loose area of white matter in T2FLAIR image). The preprocessing of skull dissection, using Otsu threshold method and morphology method, can effectively peel off non-brain tissue which interferes with white matter segmentation. In the experiment, this method is more accurate than the traditional BSE skull stripping method in Jaccard similarity measure and Dice similarity measure, compared with manual fine segmentation gold standard. Firstly, the initial contour of the brain white matter loose area is determined by the method of Otsu threshold, and then an improved C-V model is used to iterate the evolution of the initial contour. The experimental results show that the proposed method performs better than the level set method without initialization proposed by Li in the segmentation effect and the segmentation time is better. Finally, using the relationship between the eigenvalue of Hessian matrix and the direction of gray change of deep medullary vein in brain, a filter was designed to extract the main trunk of deep medullary vein in brain. Combined with morphological post-processing to disconnect the trunk and eliminate the pseudo-trunk, the result is close to the gold standard of manual fine segmentation.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：R742.89

【参考文献】