脑神经图像处理方法研究

发布时间：2018-03-24 07:35

本文选题：反卷积　切入点：多尺度　出处：《哈尔滨工程大学》2016年硕士论文

【摘要】：神经元是构成神经系统结构和功能的基本单元,神经元形态结构的三维重建对于神经系统结构和功能的研究具有重要的意义。共聚焦显微成像在成像过程中有图像退化和模糊现象。此外,神经元图像轴突和树突等树形结构具有密度、尺度变化大,像素分布不均匀和背景目标混杂等问题,这给神经元树形结构的分割带来了很大的困难,从而不利于神经元形态结构的三维重建。针对上述问题,本文将从图像复原和去燥、树形结构增强与二值化、树形结构追踪方面展开研究,主要研究内容如下:首先,针对共聚焦显微成像过程中的图像退化和模糊现象,本文引入了反卷积算法对图像进行复原处理。通过分析成像过程中的图像退化原因,建立了退化模型,并通过对比实验分析了约束最小二乘反卷积算法和Lucy-Richardson反卷积算法。此外,针对神经元图像的噪声特点和神经元的结构特点,通过对比实验分析了几种具有边缘保留特性的去噪算法对于神经元图像去噪的适应性。其次,针对神经元图像像素分布不均匀、线形结构尺度变化大的特点,本文在借鉴血管、视网膜等图像线形结构分割算法的基础上,将多尺度线形结构增强滤波运用于神经元图像轴突和树突的分割。针对分割结果存在尺度间不连续的问题,引入了各向异性高斯平滑算法,借鉴各向异性滤波的思想,本文设计了一种各向异性均值和中值滤波算法。经实验验证,这种算法在保持线形结构原始尺度不变的同时,很好的解决了多尺度融合时尺度间不连续的问题。然后,针对神经元图像背景和目标的特点,分析了传统阈值化方法对神经元图像分割的优缺点。在此基础上,本文设计了一种基于连通域的二值图像线形结构分割方法,在本文算法中,参与阈值化的对象不再是孤立的像素点,而是连接起来的连通域。经实验验证,本文算法可以较好的提取出经局部阈值化方法处理后的图像中的线形结构,本文算法为噪声干扰严重的二值图像中线形结构的提取提供了一种新的思路。最后,因为上述的研究都是以全局处理为结果的,研究人员无法对感兴趣结构或重要结构进行重点处理,也就是说缺乏交互性。此外,全局处理通常只有在图像成像质量高的情况下才能取得良好的分割效果。因此,相对于全局处理,对基于局部搜索的树形结构追踪算法的研究很有必要。本文分别从二维和三维角度研究了基于局部搜索的追踪算法和实现原理,并通过现有工具进行了神经元树形结构追踪实验分析。
[Abstract]:Neurons are the basic units that make up the structure and function of the nervous system, Three-dimensional reconstruction of neuronal morphology is of great significance in the study of nervous system structure and function. Confocal microscopic imaging has image degradation and blur in the imaging process. The dendritic structure of neuronal image has many problems, such as density, large scale change, uneven pixel distribution and mixed background target, which brings great difficulties to the segmentation of neuronal tree structure. In view of the above problems, this paper will focus on image restoration and dryness, tree structure enhancement and binarization, tree structure tracking. The main research contents are as follows: first, Aiming at image degradation and blur in confocal microscopy, this paper introduces deconvolution algorithm to restore the image. By analyzing the causes of image degradation in the imaging process, a degradation model is established. The constrained least square deconvolution algorithm and the Lucy-Richardson deconvolution algorithm are analyzed by contrast experiments. Through comparative experiments, the adaptability of several denoising algorithms with edge reservation to neural image denoising is analyzed. Secondly, aiming at the characteristics of uneven pixel distribution of neuron image and large scale change of linear structure, this paper uses blood vessels as reference. Based on the linear structure segmentation algorithm of retinal image, the multi-scale linear structure enhancement filter is applied to the segmentation of neuronal image axons and dendrites. The anisotropic Gao Si smoothing algorithm is introduced, and an anisotropic mean and median filtering algorithm is designed based on the idea of anisotropic filtering. The experimental results show that this algorithm keeps the original scale of the linear structure unchanged at the same time. The problem of discontinuity between scales in multiscale fusion is well solved. Secondly, the advantages and disadvantages of the traditional thresholding method for neuronal image segmentation are analyzed according to the characteristics of the background and target of neuron image. In this paper, a linear structure segmentation method of binary image based on connected domain is designed. In this algorithm, the object involved in thresholding is not isolated pixel, but connected connected domain. The algorithm in this paper can extract the linear structure of the image processed by the local threshold method. This algorithm provides a new way for the extraction of the linear structure in the noisy binary image. Finally, Because all of the above studies are based on global processing, researchers cannot focus on structures of interest or important structures, that is to say, lack of interactivity. Global processing usually achieves good segmentation results only when the image quality is high. It is necessary to study the tree structure tracking algorithm based on local search. The tracking experiment of neuron tree structure is carried out with the existing tools.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：R338;TP391.41

【参考文献】