CASIA OpenIR  > 毕业生  > 博士学位论文
基于体电镜的器官生理病理超微结构自动重建分析方法研究
江熠
Subtype博士
Thesis Advisor韩华
2022-05-25
Degree Grantor中国科学院自动化研究所
Place of Conferral中国科学院自动化研究所
Degree Name工学博士
Degree Discipline模式识别与智能系统
Keyword体电子显微镜 人工智能 超微结构 三维重建 分析
Abstract

生物器官生理病理超微结构的研究,对理解器官的正常生理活动,探究器官病理状态的结构差异和辅助早期疾病诊断,以及助力于新药物的研发都有着重要的意义。随着体电子显微镜成像技术和人工智能数据处理技术的快速发展,使得高精度、快速地大尺度重建器官的三维超微结构具备高度的可行性和可操作性。这能够使研究者集中力量进行器官生理和病理状态下超微结构比较解剖学研究,以期获得对疾病的发病机理更加深入的理解。

本文基于体电子显微镜的成像技术和人工智能的大规模数据处理技术,在与脑科学、细胞生物学、神经病理学的交叉领域开展生物器官生理病理超微结构的自动重建分析方法研究。首先,针对脑科学中的微观脑图谱构建,探究了小鼠视交叉上核神经元细胞核的重建方法;其次,针对细胞生物学中细胞及亚细胞结构与功能的探究,开展正常生理小鼠肝脏超微结构的研究;最后,针对神经病理学中阿尔茨海默症的研究,探究了阿尔茨海默症大鼠前额叶皮层中超微结构的变化。论文的主要成果和贡献如下:

1. 针对小鼠视交叉上核中神经元细胞核的重建问题,本文提出一种基于编解码结构的深度学习分割算法。首先,通过使用改进的残差网络和空洞卷积的空间金字塔池化模块,编码多尺度上下文信息提取图像的高维特征。然后,融合多层级的特征,以准确得到细胞核的精细边界。另外,在本算法中采用Focal loss作为损失函数解决训练过程中正负样本不均衡的问题。我们提出的分割算法在没有任何后处理的情况下,在神经元细胞核分割任务上的性能优于两个基准方法U-Net和Deeplabv3+。我们通过该算法能够自动快速地重建约265GB的体电镜数据中的所有神经元细胞核,进而可以确定神经元的数量和分布情况,这有助于微观脑图谱的构建。

2. 为了能够在纳米分辨率水平三维表征器官的亚细胞结构和探究亚细胞结构间的相互作用关系,本文使用先进的体电子显微镜对正常生理小鼠的肝脏组织进行高分辨率成像,然后设计了一个基于卷积神经网络的多细胞器分割算法,自动重建肝脏细胞中的超微结构,实现了对肝脏中多达六种细胞器的重建和量化分析。三维重建结果证实了肝脏细胞中内质网主要呈扁平的片状,并在细胞内与其他细胞器紧密结合。进一步的分析结果显示,内质网比其他细胞器具有更小的体积-膜表面积比,这表明内质网的功能可能需要较大的膜表面积。此外,精细的三维超微结构数据还揭示了内质网-线粒体的接触极其丰富,尤其是与有分支的线粒体。本研究重建了肝脏细胞中多种细胞器的精细三维超微结构,并分析了内质网与其他细胞器的相互作用关系,有望为肝脏的生理病理机制研究提供重要的结构基础。

3. 在阿尔茨海默症中,神经元和突触结构的改变被认为是认知功能障碍的主要神经生物学特征,但是由于缺乏定量分析的工具,与阿尔茨海默症有关的神经元和突触结构的具体变化仍然难以发现。为了能够定量三维表征与阿尔茨海默症有关的突触和树突等超微结构的具体变化,本文首先采用先进的体电子显微镜成像技术,获取阿尔茨海默症和其同窝野生型大鼠前额叶皮层的高分辨率序列图像。然后针对体电镜数据强各向异性的特点,设计了基于反向注意力的小尺度突触自动重建算法,和基于残差3D U-Net和图割模型的神经突起自动重建算法,以便快速准确地获取突触和树突的三维超微结构。我们从突触和树突的三维表征结果中发现,相比同窝野生型大鼠,阿尔茨海默症大鼠前额叶皮层中突触密度降低,突触并置面面积和突触后致密带体积减小。另外,发现阿尔茨海默症大鼠前额叶皮层中树突的树突棘增多,而形成突触的树突棘的比例减少,但是在树突轴上形成的突触增加。该研究结果有望为阿尔茨海默症的研究提供有价值的信息。

Other Abstract

The study of the ultrastructure of the physiological and pathological states of biological organs is critical for comprehending the normal physiological activities of the organs, exploring the structural differences between pathological and physiological conditions of the organs, and assisting in early disease diagnosis, as well as helping the research and development of new drugs. With the rapid advancement of imaging technology of volume electron microscopy and data processing technology of artificial intelligence, it is extremely conceivable and practical to reconstruct the three-dimensional ultrastructure on a large scale of organs with high precision and rapidly. It is feasible to concentrate on comparative anatomical studies of organs in healthy and diseased stages to understand disease pathophysiology better.

Based on volume electron microscopy and artificial intelligence, this thesis investigates methods on automated reconstruction and analysis of the ultrastructure of the biological organ in physiological and pathological states in the interdisciplinary fields of brain science, cell biology, and neuropathology. Firstly, we investigated the method for reconstructing the nucleus in the mouse suprachiasmatic nucleus to assist in constructing microscopic brain maps in brain science. Next, we explored the ultrastructure of the liver in physiological mice to better understand the structure and function of the cell and subcellular in cell biology. Finally, we examined the ultrastructural changes in the prefrontal cortex of Alzheimer's disease rats to analyze the disease in neuropathology.

The main achievements and contributions of this thesis are summarized as follows:

1. This thesis proposed a deep learning segmentation algorithm based on the encoder-decoder structure for automatically reconstructing neuron nuclei in the mouse suprachiasmatic nucleus. Firstly, the high-dimensional features of images are extracted by encoding multi-scale contextual information using a modified residual network and an atrous spatial pyramid pooling module. Then, the features of multiple layers are combined to accurately obtain the fine boundary of the nucleus. In addition, Focal loss is utilized as the loss function in this algorithm to overcome the problem of unbalanced positive and negative samples in the training phase. The proposed segmentation algorithm outperforms two baseline methods, U-Net and Deeplabv3+, on neuron nucleus segmentation without any post-processing. The algorithm can automatically and rapidly reconstruct all neuron nuclei from 265GB of volume electron microscopy data and then get the number and distribution of neurons, which is helpful for the construction of microscopic brain maps.

2. To three-dimensionally characterize the subcellular structure of organs at the nanoscale and investigate the interaction between subcellular components, in this thesis, we imaged the liver tissue of a physiological mouse using advanced volume electron microscopy. Then, we designed a multi-organelle segmentation algorithm based on convolutional neural networks for automatically three-dimensional reconstructing the ultrastructure of liver cells. As a result, we were able to quantify the structure of up to six organelles in the liver. The results of three-dimensional reconstruction confirmed that the endoplasmic reticulum in liver cells was predominantly flat and tightly integrated with other organelles within the cells. Further analysis showed that the endoplasmic reticulum has a lower volume-to-membrane surface area ratio than other organelles, implying that its function may require a greater membrane surface area. Additionally, three-dimensional ultrastructural data revealed an abundance of ER-mitochondria interactions, particularly with branched mitochondria. This study elucidates the fine three-dimensional ultrastructure of various organelles in the liver and the interaction between the endoplasmic reticulum and other organelles, which is expected to provide a critical structural foundation for studying physiological and pathological mechanisms of the liver.

3. Alterations of the neuronal and synaptic structure are the primary neurobiological features underlying cognitive dysfunction in Alzheimer's disease. However, due to a lack of quantitative tools, precise changes in the neuronal and synaptic structure related to Alzheimer's disease remain elusive. Therefore, to quantitatively three-dimensionally characterize the specific changes in ultrastructure such as synapses and dendrites associated with Alzheimer's disease, In this thesis, the prefrontal cortex of Alzheimer's disease and littermate wild-type rats was imaged with high resolution using advanced volume electron microscopy techniques. Then, due to highly anisotropic volume electron microscopy data, we developed a small-scale synapse automatic reconstruction algorithm based on reverse attention and designed a neurite automatic reconstruction algorithm based on residual 3D U-Net and graph cut models to quickly and accurately obtain the three-dimensional ultrastructure of synapses and dendrites. The three-dimensional characterization of synapses and dendrites revealed that synaptic density, the area of synaptic apposition, and the volume of postsynaptic density decreased in the prefrontal cortex of Alzheimer's disease rats. In addition, the three-dimensional structure of dendrites showed an increase in dendritic spines but a decrease in the fraction of dendritic spines that formed synapses in Alzheimer's disease rats. Furthermore, the number of synapses formed on dendritic shafts is enhanced in Alzheimer's disease rats. The findings are expected to provide valuable information for Alzheimer's disease research.

MOST Discipline Catalogue工学::控制科学与工程
Pages138
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/48952
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
江熠. 基于体电镜的器官生理病理超微结构自动重建分析方法研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2022.
Files in This Item:
File Name/Size DocType Version Access License
基于体电镜的器官生理病理超微结构自动重建(133007KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[江熠]'s Articles
Baidu academic
Similar articles in Baidu academic
[江熠]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[江熠]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.