【作者】 巫小蓉；

【导师】 吴效明；

【作者基本信息】 华南理工大学 ， 计算机应用技术， 2009， 博士

【摘要】 现代计算机技术为结构生物学在原子和亚细胞空间尺度上结构信息的表述、存储、获取和分析作出了巨大贡献。结构生物学是生命科学的重要学科之一。生物大分子三维结构作为其核心内容,是研究生物大分子结构及其功能相互关系的前提和基础,测定生物大分子的高分辨率结构对于了解其相应的功能具有重要意义。冷冻电镜图像生物大分子三维重构技术是获得生物大分子结构的主要方法之一。在冷冻电镜图像三维重构过程中,面临三大关键问题:(1)冷冻电镜生物大分子图像分割问题。冷冻电镜图像分割对于冷冻电镜图像高层次的特征提取、目标识别和图像配准等至关重要。但由于其低对比度和低信噪比等特点,很难达到理想的分割效果。(2)冷冻电镜图像生物大分子颗粒识别问题。生物大分子颗粒识别是冷冻电镜单粒子三维重构方法非常重要的环节,直接影响三维重构效率和三维结构精度。从低对比度和低信噪比的冷冻电镜图像中提取成千上万个生物大分子颗粒是一件相当耗时的工作,已经成为三维重构过程的瓶颈,自动颗粒识别方法的研究已成为当前关注的焦点。(3)电子断层扫描旋转图像配准问题。在电子断层扫描图像拍摄过程中,由于样本转动轴在旋转时的机械误差导致了旋转过程中样本的移动,直接影响三维结构的精度。因此,电子断层扫描旋转图像配准是电子断层扫描三维重构的关键环节。本论文围绕如何利用现代计算机技术以及数字图像处理技术提高冷冻电镜图像三维重构的效率和三维的结构精度开展了深入分析和系统研究,取得了如下的研究成果:1.针对冷冻电镜生物大分子图像的低信噪比和低对比度等特点,提出了基于小波变换和高斯差分的冷冻电镜生物大分子图像分割算法。该算法运用小波变换对冷冻电镜图像进行多尺度分解后,对低分辨率图像计算高斯差分图像,并对其进行基于灰度梯度的阈值分割,最后基于梯度信息进一步将低分辨率图像分割结果逐层过渡到高分辨率原图像上。该算法可以有效地抑制噪声,区分不同尺度的边缘信号,与canny算法以及最大类间方差方法的实验结果对比,该算法在主观评价和客观定量分析上均表现出更好的分割效果。2.本文针对冷冻电镜图像颗粒区域难以提取的问题,提出了基于高斯差分掩模的形状特征提取算法。该算法通过计算颗粒样本的高斯差分图像掩膜,提取目标对象的近似区域进行形状特征提取。通过实验与基于方差图像的方法相比,该算法所提取的形状特征的单一识别率更高,具有更好的鉴别能力和分类性能。3.针对冷冻电镜图像质量差、颗粒形状复杂多样的特点,提出了基于偏最小二乘的冷冻电镜图像颗粒识别二步方法。该二步方法包括基于高斯函数的候选颗粒检测和基于偏最小二乘的候选颗粒识别两个步骤。该方法运用基于PLS的识别机制,具有小样本学习特点,只需少量样本,就能保证在保持良好的正选率情况下,错选率很低,实验结果表明,该方法对方形颗粒的错选率为14%,对圆形颗粒的错选率只有7%。其次,采用了多种类型特征集合,可以对不同颗粒形状进行识别。实验结果验证了该方法具有通用性强,识别率高等特点。与同类算法相比,该方法在错选率方面具有优势。4.针对电子断层扫描图像受噪声污染严重的特点,提出了基于局部互相关的电子断层扫描旋转图像配准优化模型。该模型克服了传统全局互相关方法固有的对噪声敏感的局限性,使配准结果更精确。该模型还具有良好的可扩展性,即可在此模型基础上,灵活定义相似度评价函数而无需改变整个算法框架。实验结果显示,该方法比同类基于全局互相关的方法表现出更好的配准精确度。5.为了获得优化模型的最优解,需要对整个变换系数空间进行搜索,使算法效率非常低。根据图像序列前后图像偏移的相关性,提出了电子断层扫描旋转图像配准优化模型的自适应搜索机制。该机制能缩小变换方程系数空间的搜索范围,使配准过程能快速有效的进行。实验数据显示,使用该搜索机制,算法效率至少可以提高70%。更多还原

【Abstract】 Mordern computer technology has made a significant contribution to the study of structural biology on expressing, storing, acquiring and analyzing structural information. Structural biology is one of the most important subjects of life sciences. Its key research content, the three-dimensional (3D) structure of biological macromolecules, is the premise and foundation of studying the relationship between macromolecule structure and function. Determining high resolution 3D structure of macromolecule is significant to understanding its corresponding function. Electron cryomicroscopy (Cryo-EM) is one of the main methods to obtain the macromolecules structure.There are three key issues which is crucial to Cryo-EM 3D reconstruction. The first problem is segmentation of Cryo-EM image, which is crucial to the high-level processing of electron micrograph, including feature extraction, particle recognition, image registration and so on. But Cryo-EM image has the characteristic of low contrast and low signal-to-noise ratio (SNR) and is difficult to acquire ideal segmentation effect. The second is macromolecule particle recognition from Cryo-EM image, which is a crucial step of single particle reconstruction and has directly effect on the efficiency of Cryo-EM 3D reconstruction and the resolution of 3D structure. Selecting hundreds of thousands of particles from low contrast and low SNR Cryo-EM image is one of the major bottle-necks in advancing toward achieving atomic resolution reconstruction of biological macromolecules. The third problem is alignment of transmission electron microscope tilting series. Mechanical inaccuracies of the specimen holder cause specimen movement during tilting, which requires appropriate compensating adjustment. Accurate image alignment is needed for computing three-dimensional reconstructions from transmission electron microscope (TEM) tilt series.This paper has made deeply research on how to use modern computer technology and image processing technology to enhance the efficiency of CryoEM three-dimensional reconstruction and the resolution of 3D structure. Several finding has been made, which is as follows.1. This paper provides a new Cryo-EM image segmentation method based on wavelet transform and difference of Gaussian (DoG), which performs multi-scale decomposition based on wavelet transform at first. And then this method gets the DoG image of low-resolution image and does threshing segmentation based on the combination of gray and gradient information. At last, the segmentation result in low-resolution image is transferred to the high-resolution image based on the gradient. This method can overcome the flaw of the low SNR, and experiments show that it can get better segmentation effect than the canny algorithm and OTSU method.2. It is difficult to get particle region in Cryo-EM images, which leads to the difficulty of shape fearture extraction. This paper proposed a shape feature extraction algorithm based on DoG mask. This algorithm only obtains the approximate region of objects, from which the shape features are extracted. Experiments show the shape feature extracted by this method has better ability of classification than the method based on variance image.3. As to the problem of particle recognition, a two step approach for the particle recognition from Cryo-EM images has been developed in this paper, which is based on the principle of partial least square (PLS). This approach involves two steps: the detection of candidate particles based on Gaussian function and the recognition from these candidates based on PLS. This method is characterized by small sample study and can obtain satisfied recognition result with a few samples needed. Experiments show it can get low error-select-rate, which is 14% while selecting rectangular particle and only 7% while selecting circle particle. Also, this method uses a wide range of feature set to recognize particles, which is suitable to different type of particles and different image qualities. The experiment results show it has advantage on the generability.4. As to the problem of tilting alignment, a TEM tilting alignment optimization model based on local cross-correlation is established in this paper. This model can do well in noisy TEM images. And it also has good extensibility, which means it can change the similarity evaluation function freely without the change of the model. Experiments show that this method has more accurate aligning result than the global cross-correlation method.5. To obtain the optimum, it is needed to search the whole transform coefficients space, which is time-consuming. This paper proposes an adaptive searching mechanism according to the relativity between two consecutive tilt images, which can greatly reduce searching range. Experiment shows that it can enhance efficiency by 70% at least.更多还原