CASIA OpenIR  > 毕业生  > 硕士学位论文
 基于深度学习的神经组织微观结构重建算法研究 饶强 Subtype 工学硕士 Thesis Advisor 韩华 2017-05-22 Degree Grantor 中国科学院研究生院 Place of Conferral 北京 Degree Discipline 模式识别与智能系统 Keyword 神经组织微观结构重建 机器学习 深度卷积神经网络 形状群相似active Contour Kcf Abstract 神经组织微观结构重建是探索和理解神经功能的重要环节，对于神经回路结构解析、神经性疾病机理探索、神经功能和结构映射等研究具有重要的意义。因此，基于神经电镜图像进行大规模神经回路的重建吸引了越来越多研究者的兴趣。本文基于序列电镜图像研究神经结构重建的自动化算法。为了满足神经结构重建对精度和速度的要求，针对神经电镜图像具有歧义、结构信息复杂、图像上目标自动识别分割困难等特点和难点，本文围绕图像像素级识别、分割、序列图像上目标跟踪等技术，采用了深度学习和KCF等机器学习算法展开研究。本文的主要研究内容和创新点归纳为如下：   一、本文对神经元分割问题的研究发现神经膜结构的检测是完成神经元分割的关键。针对神经元膜结构的检测问题，本文提出基于深度卷积神经网络对神经膜结构进行检测；首先用基于逐像素识别的方法，构造了以像素周围图像块上下文信息为输入的卷积神经网络模型，用以获取神经膜识别概率图（Membrane Detection Probability Map, MDPM）；在进一步研究的基础上，本文基于全卷积网络（Fully Convolutional Network, FCN）提出了一种融合空间多尺度卷积特征和深度上下文特征的分割网络SPPUNet，大大的提升了生成MDPM的速度并改善了精度。   二、在深度卷积网络获取MDPM的基础上，对于神经元分割和重建问题，本文提出了首先用多尺度中值滤波预先处理MDPM，对经过预处理后的MDPM应用双尺度控制种子点分水岭算法进行分割的神经元分割流程，完成了果蝇蘑菇体1693片$14k\times 14k$ 扫描电镜图像神经元的分割。本文还基于序列分割图像利用启发式的神经连接追踪算法完成了$10\mu m\times 10\mu m \times 10 \mu m$ 体量神经结构的自动重建。   三、进一步，本文针对果蝇蘑菇体电镜图像的特点，改进了启发式神经连接追踪的方法，提出了基于形状群相似约束的Active Contour成组分割神经元的和启发式与KCF结合追踪单根神经元的算法，并把形状渐变信息以加权的方式引入成组分割中，最后提出了一种联合分割追踪算法框架。 Other Abstract Neuronal microstructural reconstruction is central for functionally exploring and understanding neural networks in biology, and is of significance for the studies of  topology of neural networks, nerve diseases and mapping between neuronal structure and it's functions. Large scale neural networks reconstruction has recently garnered general interest. we study the automated neuronal reconstruction method based on the sequence of neural scanning electron microscopy(SEM) images. To fulfill the requirements of neuronal reconstruction for accuracy and speed, taking into consideration of the ambiguity, complexity, and  difficulties of automated recognition and segmentation microstructure of the neural SEM images, we investigated the imagery pixel-wise recognition, segmentation and object tracking by using machine learning methods including deep learning and KCF.  the main works and creations of this dissertation are listed as follows:   1) We found that the membrane detection is the key of neuron segmentation in our studies. To solve the problem of neuron membrane detection, we proposed to use deep convolutional neural networks (DCNN). We firstly constructed the pixel classifier based on DCNN with the image block centred around each pixel as its input. With the classifier, we can get the membrane detection probability map(MDPM) .  By furher studies, based on the fully convolutional neural network (FCN), we proposed a network named as SPPUNet that fuses the multiply spatial scale convolutional features and deep contexture features. SPPUNet significantly improved the speed of generate MDPM and also improved the performance to some degree.   2) After getting the MDPM, to solve the problem of neuron segmentation and reconstruction, we proposed a pipeline that firstly  preprocessing the MDPM with multi-scale median filter to filter out the blur noises, and then applying the double scale marker controlled watershed on the preprocessed MDPM. Based on the pipeline, we got 1693 sections, $14k\times 14k$ fully segmentation of the fly mushroom body neuronal SEM images. To reconstruct the neurons from the segmentation maps, we proposed the heuristic tracking and linking method, based on which we densely reconstructed neurons within $10\mu m\times 10\mu m\times 10\mu m$ volume. 3) Further more, considering characteristic of the fly mushroom body SEM images, we improved the heuristic tracking and linking method.  We proposed the method that group shapes similarities active contour segmenting neurons in group of the consecutive sections and combining heuristic  linking with KCF tracking for tracing single neurons. After all, we proposed a framework that segmenting and tracing neurons simultaneously. Subject Area 模式识别与智能系统 Language 中文 Document Type 学位论文 Identifier http://ir.ia.ac.cn/handle/173211/14646 Collection 毕业生_硕士学位论文 Affiliation 中国科学院自动化研究所 Recommended CitationGB/T 7714 饶强. 基于深度学习的神经组织微观结构重建算法研究[D]. 北京. 中国科学院研究生院,2017.
 Files in This Item: File Name/Size DocType Version Access License Thesis.pdf（8992KB） 学位论文 暂不开放 CC BY-NC-SA Application Full Text
 Related Services Recommend this item Bookmark Usage statistics Export to Endnote Google Scholar Similar articles in Google Scholar [饶强]'s Articles Baidu academic Similar articles in Baidu academic [饶强]'s Articles Bing Scholar Similar articles in Bing Scholar [饶强]'s Articles Terms of Use No data! Social Bookmark/Share