基于细胞识别和事件探测算法对钙成像数据的自动化分析研究及应用

发布时间：2018-05-07 05:10

本文选题：双光子钙成像 + 细胞识别与分割　；参考：《西南大学》2017年硕士论文

【摘要】：神经科学是一门重点研究脑科学的综合性学科。在最近20年期间,神经科学经历着飞速的发展,对类脑人工智能的进步及各种神经及精神类疾病的治疗有着非常重大的意义。其中计算机技术对于推动神经科学的发展起着不可替代的作用。随着神经活动记录技术的发展与新的纪录方法的出现,脑科学数据的分析成为一重大难题,急需借助计算机自动化分析技术来取代手工分析,提高分析的效率与精度。在过去的10年中,双光子钙成像技术已经被广泛地应用于神经元群的功能活动成像,并且可以很容易地与细胞类型特异标记物结合用于分析特定类型神经元环路的功能。为了达到这一目标,就需要在单细胞水平上进行神经活动的分析。然而,人工方式进行细胞的识别费时并且标准很难统一。因此,通过计算机技术来自动地、精确地快速识别单个神经细胞的位置和轮廓具有重要价值。在此基础上,通过提取单个神经细胞荧光强度变化来分析动作电位相关的活动,以此和行为学变量联合可解析大脑特定皮层区域的工作机制。可见,神经细胞的识别分割和单个细胞钙信号事件的探测是光学脑功能成像数据分析工作中的基础并具有至关重要的作用。故本文研究工作主要分为两个部分:(1)提出一种新的细胞识别与分割方法。该方法主要分为3个步骤:(a)对钙成像数据的细胞图像,利用多尺度拉普拉斯高斯滤波(Multi_LoG)定位局部极值从而实现对细胞的种子点(中心点)的初步探测;(b)利用卷积神经网络(CNN)算法进行细胞的进一步判别,降低探测结果的假阳性;(c)利用TWANG算法对细胞进行边缘检测,该算法的优点是具有精确分割能力的同时计算复杂度低,从而可以进行细胞边缘的快速分割。本文将此方法应用于开源细胞图像(benchmark)和来自第三军医大学脑研究中心的双光子钙成像细胞图像,并与一些已发表论文中的细胞识别分割算法比较。(2)对于钙成像数据的荧光亮度变化曲线,提出一种新的钙事件探测方法。首先通过分析建立反应动作电位活动的钙事件的相关特征参数,利用一个滑动的基线窗获得噪声水平的估计值。然后再提取基线相邻的探测窗口数据中钙事件的特征信息,并且与多项特征参数进行匹配以判断该该事件活动是否满足钙事件的特征条件。在成功判断的基础上,进一步提取钙事件的初始点,峰值点和结束点的位置信息。最后在钙事件数据中,将已成功探测的钙事件的噪声水平进行再次估计用作下一步的探测。以此完成一个钙事件的探测过程,并通过这种方式循环直至整个钙成像荧光数据的事件探测完成。本文将此方法应用于仿真钙事件数据和来自第三军医大学脑研究中心的双光子钙成像荧光亮度曲线,并与现已发表论文中的一些钙事件探测算法比较。基于以上工作,对本文分析方法进行评估的方案主要是比较这些方法所得到结果的召回率(Recall),精确率(Precision)和F分数(F-score)这三个参数的值。通过应用于开源和仿真数据,证明本文方法的有效性和正确性;通过应用于真实数据,证明本文方法的实用性和可行性。借助对上述多项数据的应用,将对比方法与本文提出方法产生结果的平均值进行比较并对其进行显著性差异分析,证明本文方法可在自动化的分析过程中获得更精确的神经细胞的识别分割效果和提高其对应的钙事件探测精度,这对于将来大规模应用在双光子成像数据分析中具有重要作用。
[Abstract]:Neuroscience is a comprehensive subject which focuses on the research of brain science. During the last 20 years, neuroscience has developed rapidly. It has great significance for the progress of brain artificial intelligence and the treatment of various neurologic and mental diseases. With the development of neural activity recording technology and the emergence of new record methods, the analysis of brain science data has become a major problem. It is urgent to replace manual analysis with computer automation analysis technology to improve the efficiency and accuracy of analysis. In the past 10 years, the two-photon calcium imaging technology has been widely used in the group of neurons. Functional activity imaging, and can easily be combined with cell type specific markers to analyze the function of a specific type of neuron loop. In order to achieve this goal, the analysis of neural activity is needed at a single cell level. However, the artificial way of cell recognition is time-consuming and the standard is difficult to unify. Therefore, through Computer technology is of great value to automatically, accurately and quickly identify the location and contour of a single nerve cell. On this basis, the action potential related activities are analyzed by extracting the changes of the fluorescence intensity of a single nerve cell, which can be combined with the behavioral variables to analyze the working mechanism of the specific cortical regions of the brain. The recognition and segmentation of cells and the detection of single cell calcium signal events are the basic and important role in the analysis of optical brain functional imaging data. Therefore, the main research work of this paper is divided into two parts: (1) a new method of cell recognition and segmentation is proposed. This method is divided into 3 steps: (a) fine calcium imaging data Cell image, using the multi-scale Laplasse Gauss filter (Multi_LoG) to locate the local extremum to realize the preliminary detection of the seed point (center point) of the cell; (b) use the convolution neural network (CNN) algorithm to further distinguish the cells and reduce the false positive of the detection results; (c) the edge detection of the cells by the TWANG algorithm, the algorithm This method is applied to open source cell image (benchmark) and the image of two-photon calcium imaging cells from the Third Military Medical University brain research center in this paper, and is compared with the cell recognition and segmentation algorithms in some published papers. (2) a new method of calcium event detection is proposed for the luminance change curve of the calcium imaging data. First, by analyzing the related characteristic parameters of the calcium event of the action potential activity, a sliding baseline window is used to obtain the estimated value of the noise level, and then the calcium event in the detection window data adjacent to the baseline is extracted. The feature information is matched with multiple characteristic parameters to determine whether the event activity meets the characteristics of the calcium event. On the basis of the successful judgment, the initial point, the peak point and the end point of the calcium event are further extracted. Finally, the noise level of the calcium event has been successfully detected in the calcium event data. This method is used to simulate calcium event data and two photon calcium imaging luminance curves from the heart of the Third Military Medical University, and the method is applied to the simulation of calcium event data and the luminance curve of the two photon calcium imaging from the heart of the Third Military Medical University. Compared with some of the calcium event detection algorithms in the published papers, based on the above work, the main evaluation of this method is to compare the recall rate (Recall) of the results obtained by these methods, the value of the three parameters of the accuracy rate (Precision) and the F fraction (F-score). The validity and correctness of the method are proved, and the practicability and feasibility of this method are proved by application of the real data. With the use of the above data, the comparison method is compared with the average value produced by the method presented in this paper, and the significant difference analysis is carried out. It is proved that this method can be obtained in the automated analysis process. The recognition and segmentation of more accurate neural cells and the improvement of its corresponding detection precision of calcium events are important for large-scale application in the future analysis of two-photon imaging data.

【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R318;TP391.41

【参考文献】