【作者】 李小兵；

【导师】 邱天爽； Su RUAN； Stéphane LEBONVALLET；

【作者基本信息】 大连理工大学 ， 信号与信息处理， 2010， 博士

【摘要】 本论文的研究工作源于中法联合培养博士研究生项目。所研究的内容主要涉及两方面：磁共振(MRI)图像脑肿瘤分割和脑电信号(EEG)脑癫痫检测。首先,磁共振图像脑肿瘤分割方面的研究工作主要集中于以下几个方面。准确的脑肿瘤组织分割对脑肿瘤以及其他大脑疾病的诊断和治疗十分重要,通过对脑肿瘤的分割和跟踪,医生可以测量肿瘤的各个参数,例如肿瘤的大小和位置等,进而可以分析和定量评估脑肿瘤状态和生长变化过程,评价治疗的效果等等。准确、通用的脑肿瘤自动分割方法的研究又是一项十分具有挑战性的工作。因为,一般情况下MRI图像中的脑肿瘤边缘模糊,同时所要分割的肿瘤在MRI图像中存在很大差别。基于以上认识,本论文的目的是研究设计一个准确、通用的脑肿瘤自动分割系统,对特定脑肿瘤患者的治疗过程进行跟踪。在治疗期间,患者每四个月进行一次MRI扫描,扫描的图像经过分割和定量分析,得到真实临床数据,帮助医生对肿瘤的形态进行跟踪分析,用以评价治疗的效果。论文围绕以上设计目的展开,在回顾医学图像分割技术尤其是水平集方法的基础上,结合大脑MRI图像本身的特点,运用大脑中矢面估计算法、脑肿瘤初始轮廓的搜索算法等技术,设计完成脑肿瘤自动分割系统。系统经过图像预处理,肿瘤分割和分割结果比较评价三个主要阶段,完成脑肿瘤的分割工作。在图像预处理阶段,主要进行一些基本的图像操作,对图像的质量、图像的位置等内容进行改善,包括纠正图像的帧与帧之间,以及同一帧内灰度上的不一致；保证不同时间采集的图像之间空间上的基本一致性,对其进行配准。肿瘤分割阶段是系统最关键部分,它由三个步骤组成：初始帧确定、帧内分割和帧间分割。首先确定初始帧,利用大脑中矢面信息,通过对图像的对称性检测来判断是否存在潜在的脑肿瘤,结合分水岭和图像形态学等方法计算出肿瘤的大致位置和肿瘤的灰度值,并得出肿瘤的初始轮廓。其次,对初始帧进行帧内分割：利用水平集及其改进算法对肿瘤的初始轮廓进行演化,根据改进的演化停止条件,决定停止的位置,得到肿瘤的边界真实,完成帧内分割。最后进行帧间分割：将初始帧的分割结果投影到相邻的图像帧上,利用相同的帧内演化算法求得所有帧内的肿瘤边界,完成肿瘤分割的全过程。肿瘤分割和分割结果比较评价阶段,就是在治疗期间不同时间采集的所有MRI图像分割完成之后,生成肿瘤的量化数据和可视化图像,医生可以根据分割的结果,对肿瘤的生长状况进行分析、比较,评价治疗的效果,指导进一步的治疗。本论文所设计的脑肿瘤自动分割系统能够自动地分割MRI序列图像,并且通过与有经验的医生手工分割结果相对比,表明所提出的方法的结果与医生手工分割的结果具有很好的一致性,验证了本系统的有效性。同时实验显示,本系统所采用的水平集算法对所选参数并不十分敏感,因此对一般的MRI序列脑肿瘤图像均有良好的适用性。其次,脑电信号分析和脑癫痫检测方面的工作主要涉及以下几个方面。近年来,时频分析方法在脑电信号研究方面的应用发展很快,因其充分考虑了脑电信号的非平稳特性,在时频平面上研究信号的时变特性,因此可以在时间和频率上同时具有很好的分辨率。但是,在利用其进行脑电分析时,常常伴随有因不同频率交叠而生成的交叉项,从而产生被误判的虚假信号。本论文结合采用时频分析方法与奇异值分解方法,以减少交叉项的影响,并通过时频分布的差异测度方法对脑癫痫信号进行检测,取得较好的效果。为了进一步抑制交叉项,采用先对脑电信号进行经验模式分解然后重构的方法,因为重构时考虑到待检测的脑电信号的特性,可以强调期望检出的信号分量而抑制其他分量,所以能够较好地达到抑制交叉项的目的。依此方法实现的脑癫痫检测效果较好。本论文研究工作的主要贡献包括：(1)引入了大脑中矢面估计算法。正常人类的大脑的对称性也反映在轴向(横断面)核磁共振图像中,利用这种对称性,我们可以通过分析轴向核磁共振图像,来估计出中矢面的位置。这种全局信息能够为之后的局部分割提供帮助。(2)提出了脑肿瘤初始轮廓的搜索算法。在正确估计出大脑中矢面的基础上,通过计算中矢面两侧的差异,找出一个MRI序列图像中所有帧中两侧差异最大的图像帧,并且借助分水岭算法和数学形态学算法,估计出肿瘤的初始轮廓。(3)提出了一种改进的水平集主动轮廓模型。采用水平集主动轮廓模型,将肿瘤的边界定位问题转化为曲线演化问题,利用水平集方法,对边界进行细分。在C-V模型的基础上,提出了停止水平集曲线演化的条件,使其能够正确地收敛停止在肿瘤的最终边界。(4)设计了3D数据比较算法,用以评价比较分割出的肿瘤的生长变化状态。(5)提出了一种基于时频分布与奇异值分解相结合的减少时频分析交叉项的方法。(6)实现了一种基于时频分析与经验模式分解方法相结合的抑制时频分析交叉项的方法。更多还原

【Abstract】 This paper comes from the Sino-French joint training doctoral students’project. The research mainly involves two aspects:MRI brain tumor segmentation and detection of EEG epileptic brain.First, the magnetic resonance image segmentation research in brain tumor research focuses on the following aspects.Accurate and robust brain tissue segmentation is a very important issue for the diagnosis and treatment of brain tumors and the study of some brain disorders. One example is to analyze and estimate quantitatively the growth process of brain tumors, and to evaluate effects of some pharmaceutical treatments in clinic. Once a tumor is found, physicians can measure various quantities, such as the size and the location of tumors. However, tracing a tumor in 3D manually by an expert is not only exceedingly time consuming, but also exhausting for the expert leading to human errors. Therefore, it is necessary to develop segmentation tools with minimum manual intervention.Automatic, accurate and robust brain tissue, and brain tumor segmentation is a great challenging task because it usually involves a large amount of data with sometimes artifacts due to patient’s motion or limited acquisition time and soft tissue boundaries. In addition there is a large class of tumor types which have a variety of shapes and sizes, and may appear at any location and in different image intensities. Some of them may also deform the surrounding brain structures. The existence of several MR acquisition protocols can provide different information on the brain. Each image usually highlights a particular region of the tumor. Thus, automated segmentation with prior models or using prior knowledge is difficult to implement.In this context, the aim of our project is to develop a framework for an automatic, robust and accurate segmentation of a large class of brain tumors in MR images. The built system based on this framework is used to follow a specific patient in his whole therapeutic period while his MRI images acquired once every four months a year, allowing the clinician to monitor tumor developing states and evaluating the therapeutic treatment.The framework consists of three steps:image preprocessing, tumor segmentation and result comparison and therapy evaluation.Image preprocessing. In this step, operations such as:reduction of intensity inhomogeneity and inter-slice intensity variation of images, spatial registration (alignment) of the input images are performed. This section prepares images and some global information on the brain to be used in the segmentation section. Tumor segmentation. First, the approximate symmetry plane of the MRI volume is computed, and the initial contour of the tumor, if the tumor is present in the image, is searched by utilizing the symmetry plan information. Second, a level set method is used to refine the initial contour to get the tumor boundary. Last, the tumor boundary is, in the middle part of the MRI volume in general, projected to its adjacent slices for the new initial contours of the adjacent slices. The same refinement algorithm is applied to get all tumor boundaries through the whole volume. All the boundaries in the same volume are used to reconstruct 3D tumor volume for the tumor quantitative measurements.Result comparison and therapy evaluation. In this last step, by following up the tumor variations in the therapeutic period, the clinician can carry out comparison studies according to the medical requirement, and give the evaluation of the therapeutic treatment.Experimentation and validation results show that the proposed segmentation approach has the ability to segment MRI volumes automatically, and has a relatively good segmentation effect; Experimentations also show that the segmentation results are not too sensitive to the parameters in level set evolution. The built system based on this framework is used to follow a specific patient in his whole therapeutic period while his MRI images acquired once every four months a year, allowing the clinician to monitor tumor developing states and evaluating the therapeutic treatment.Second, EEG analysis and brain seizure detection work mainly involves the following aspects.In recent years, researchers have been using a variety of signal analysis and processing techniques, try to design for automatic EEG diagnostics. Time-frequency analysis of these methods are developing fast, its full consideration of the characteristics of EEG non-stationary in time-frequency plane, time-varying characteristics of the signal, it can be with good resolution in time and frequency. Although the time-frequency analysis method has these advantages, however, when using it for EEG analysis, it is often accompanied by overlapping of different frequencies generated by the cross-term, resulting in a false judgment.In this thesis, combined with time-frequency analysis method and singular value decomposition method, the impact of the cross-term is diminished, and different time-frequency distribution measurement methods are tried to detect cerebral epileptic signal to obtain good results. To further suppress cross terms, the empirical mode decomposition and reconstruction method are used, because the reconstruction takes into account the characteristics of EEG, it can be expected to inhibit the detection of the signal components of other components, so it can better achieve the purpose of cross-term suppression. This is the way to achieve better detection results of brain seizures. The main contributions of this research work discussed in the dissertation mainly include:(1) Introduction of a mid-sagittal plan estimation algorithm.As to axial MRI images, the symmetry plane of a normal brain is a good approximation of the mid-sagittal plane, best separating the hemispheres. To determine the location of the plane, we compute a degree of similarity between the slice image and its reflection with respect to a plane, by utilizing each slice and combining results from multiple slices. The best plane is then obtained by maximizing the similarity measure.(2) Proposition of an initial contour seeking algorithm.After the extraction of the mid-sagittal plane, we then calculate the differences between two hemispheres. The slice with the largest difference is checked out. Using a combination of watershed and morphology algorithms, the region without symmetry can be determined, which is considered as the initial contour of the tumor in this slice. Usually it is the largest size contour of the tumor.(3) Proposition of an improved level set formulation based on active contour model.To refine the initial contour obtained in the above step, which is not accurate enough, we use edge information. An improved level set formulation based on active contour model is applied for this purpose. The proposed method tries to combine region and edge information, thus taking advantage of both approaches while cancelling their drawbacks.(4) Implement software to perform 3D data comparison.After all the tumor data of the volumes in the therapeutic period have been segmented, a 3D reconstruction algorithm is designed to visualize the tumor and quantify the tumor information making it convenient for the clinician evaluate the therapeutic treatment.(5) Proposition of a time-frequency distribution based on singular value decomposition method of cross-term reduction approach.(6) Proposition of a time-frequency analysis based on empirical mode decomposition method of cross-term suppression approach.更多还原