【作者】 张翔；

【导师】 徐涛；

【作者基本信息】 华中科技大学 ， 生物物理学， 2017， 博士

【摘要】 随着显微镜成像技术的蓬勃发展,通过荧光标记某一个生物大分子,研究人员能够获得该生物大分子动态变化过程的图像序列,然而如何处理好这些长时间的动态图像序列是当今的一个难点和热点。研究人员经常手动分析荧光图像,这个过程既耗时又耗力,而且人工分析还会产生主观偏见。也就是说,分析结果在很大程度上取决于个人的技能、判断和偏好。因此本论文主要对生物图像处理算法进行了详细研究,尝试用它们来自动化处理这些复杂的生物图像数据,以及用它们定量描述生物过程,同时结合生物学研究来得到一些新的发现。本论文的第一部分主要研究了活细胞中运动囊泡的识别和追踪算法。在囊泡识别方面,我们尝试将自适应阈值算法、局部最大值算法、小波变换算法、分水岭算法、压缩感知算法和机器学习算法等应用到活细胞的囊泡识别中,通过实际对比分析,我们发现小波变换算法识别到的囊泡数目最多,小波变换加上分水岭算法以后,识别到的正确囊泡数目进一步增加。同时本论文还提出了一种“self-checking”算法用来校正其他算法的检测错误,以进一步提高检测的正确率。此算法的主要思想是构建一个由多核函数叠加构成的模型,然后用这个模型去拟合无法分辨时刻的数据,通过拟合后的模型与真实数据的残差及拟合得到的核函数的参数来确定该时刻囊泡的数目及各囊泡的中心位置。在囊泡追踪算法方面,我们详细研究了多假设追踪算法、卡尔曼滤波算法以及交互多模型算法,并提出了一个优化的囊泡轨迹追踪流程图。通过优化的囊泡识别和追踪算法,我们对葡萄糖刺激前后小鼠β细胞囊泡运动轨迹进行了分析。通过分析发现,葡萄糖刺激后,囊泡的轨迹数量将会增加,囊泡的平均锚定时间会减少,这是由于胰岛细胞需要借助囊泡的转运和分泌来调控血糖平衡。同时我们对加入刺激后,WT细胞和KO细胞中囊泡的锚定时间进行分析,发现Spire1 KO细胞中囊泡的平均锚定时间比WT细胞中的减少一半,这表示Spire1可能参与了囊泡锚定,并在稳定囊泡锚定中发挥了作用。总的来说,通过囊泡识别追踪算法,我们在亚细胞水平定量分析了活细胞中囊泡的活动。本论文的第二部分主要研究了单分子荧光能量共振转移技术(smFRET)在生物学中的应用。在本部分中,我们主要分析了小波变换算法和滚球算法在单分子荧光图像中去除背景噪声及提取单分子信号的性能,并且提出了自动处理单分子FRET的算法流程图。同时我们计算了 15bpDNA的FRET效率的统计直方图,通过高斯拟合得到15bp DNA的FRET效率约为0.634,对应的距离为5.47nm。该计算距离与实际距离有0.37nm的偏差,可能是由于标记染料的取向所引起的。同时我们还计算了Syntaxin1的FRET效率的统计直方图,通过高斯拟合发现在静息状态下Syntaxin1有两个不同的状态,其中一个是距离较大的“开”状态,另一个是距离较小的“闭”状态。本论文的第三部分主要研究自动对焦算法评估线虫脂滴图像。在本部分中,我们系统评估了常用的16种自动对焦算法,以确定最适合线虫脂滴荧光图像的自动对焦算法。我们发现,WT算法精度较差,计算时间较长,因此它不适合于评价线虫脂滴荧光图像。综合考虑计算时间和精度,我们认为ATEN、MDCT和TEN比较适合线虫脂滴荧光图像。其中,ATEN的精度最好。在一个自动化筛选系统中,我们经常需要较高的精度和较快的采集速度,在这种情况下,也可以先使用速度比较快的TH算法进行粗搜索,再使用ATEN算法进行细搜索。更多还原

【Abstract】 With the rapid development of microscopy imaging technology,it is easy to acquire time-lapse images of biological macromolecules’ dynamic process by fluorescent labeling them.However,how to deal with these long time image sequences is difficult and hot.Biologists often analyze fluorescence images manually.This process is time-consuming and laborious.In addition,manually analysis will produce subjective bias.In other words,the results of the analysis are largely dependent on individual skills,judgments and preferences.Therefore this paper focuses on studying biological image processing algorithms,and develops automatical analysis to handle these complex data and to fully exploit them for describing biological process on a quantitative level and building accurate mathematical models of dynamic structure to get some new biology findings.In the first part of this paper,we studied the detection and tracking algorithms of moving vesicles in living cells.In the aspect of vesicle detection,we try to apply adaptive threshold algorithm,local maximum algorithm,wavelet transform algorithm,watershed algorithm,compressive sensing algorithm and machine learning algorithm to living cell for vesicle detection.After comparing and analysis,we found that the wavelet transform method is able to detect the largest number of vesicles,and the number of correct detection can be further increased by adding watershed algorithm.Meanwhile,a"self-checking" algorithm was proposed to correct the error of other algorithm in order to improve the detection accuracy further.The main idea of this algorithm is to construct a multi-kernel function superposition model and use the model to fit the data at the indistinguishable moment;the number of vesicles and the central positions of vesicles are determined from the set based on χ2-statistics of the residuals in least-square fits of the models to the image data.In the aspect of vesicle tracking algorithm,we studied the multiple hypothesis tracking algorithm,Kalman filtering algorithm and interactive multiple model algorithm,and proposed an optimized flow chart of vesicle tracking.We analyzed the movement track of the vesicles in β cell using optimized tracking algorithm before and after glucose stimulation.We found that the number of vesicle traces increased and the average docking time of vesicles decreased after glucose stimulation based on our tracking analysis.This is because β cells will release insulin to regulate glucose balance with the help of vesicle translocation and secretion after glucose stimulation.Meanwhile we analyzed the docking time of WT cell and KO cell after stimulation.We found that the average docking time of Spire1 KO cell reduced less than half of WT cell,which indicated that Spire1 may be involved in vesicle docking,and play a role to stabilize the vesicle docking.In a word,we quantified the vesicles activity in mice β cell by tracking analysis on subcellular level.In the second part of this paper,we focused on the application of single molecule fluorescence resonance energy transfer(smFRET)in biology.In this part,we analyzed the performance of wavelet transform algorithm and rolling ball algorithm in background noise removal and signal extraction from single molecule images,and proposed the flow chart of automatic processing of single molecule FRET.Meanwhile,we calculated the statistical histogram of 15 bp DNA FRET efficiency.By Gaussian fitting,we obtained the result that 15 bp DNA FRET efficiency was approximately 0.634 and the corresponding distance was about 5.47nm.The discrepancy between the real distance and calculated distance was 0.37nm,which may be due to the orientation of the labeled dyes.In general,these two algorithms can be applied to the single molecule FRET images to improve processing efficiency.Moreover,we also calculated the statistical histogram of Syntaxinl FRET efficiency.By Gaussian fitting,we found that Syntaxinl has two states.One of them may be "on" state which has a larger distance,and the other is "closed" state which has a smaller distance.In the third part of this paper,we focus on the evaluation of autofocus algorithms for automatic detection of Caenorhabditis elegans lipid droplets.In this paper,We evaluated 16 autofocus algorithms which were collected form well-known algorithms as well as the most recently proposed focusing algorithms to found the algorithm which is best for C.elegans lipid droplets images.In our study,WT show a poor accuracy,and has a long computational time,so it is not suitable for C.elegans lipid droplets,although WT may have a good performance in other applications.Comprehensive consideration of accuracy and computational time,we recommend ATEN,MDCT and TEN for C.elegans lipid droplets.Moreover,ATEN achieves the best accuracy.In an automatic screening system,we often require both high accuracy and fast acquisition.In this case,we can apply the fastest algorithm TH for rough search,and then apply ATEN algorithm for fine search.更多还原