CASIA OpenIR  > 毕业生  > 博士学位论文
面向生物多模态显微成像数据的大规模三维重建方法及应用
刘家正
2024-05-18
Pages210
Subtype博士
Abstract

在微观尺度下重建生物组织对于深入理解组织结构、功能和疾病至关重要。传统的二维成像分析在捕捉组织的复杂三维结构方面存在局限性,而这对于阐明组织在健康和疾病中的作用至关重要。近年来,共聚焦显微镜、双光子显微镜、电子显微镜和组织透明化方法等微观组织重建技术的进步,使得研究人员能够以前所未有的分辨率和深度对组织进行三维重建和分析。这些技术为研究组织发育、形态发生和疾病发生机制提供了新的途径,同时也为药物发现和测试提供了便利。

本文基于微观成像技术和人工智能的大规模数据处理技术,在与细胞生物学、植物生物学、神经生物学及脑科学的交叉领域开展生物组织超微结构的自动重建分析方法研究。首先,针对细胞生物学中的不同品系细胞的三维结构差异,基于结构光照明显微镜和光片荧光显微镜探究了拟南芥叶片中不同生长阶段的内质网结构差异和不同品系杨树种子和油松茎尖分生组织的空间分布及三维形态;其次,针对植物生物学中细胞及亚细胞结构与功能的探究,使用连续切片自动收集系统和扫描电子显微镜,探究了不同发育时期和发育部位拟南芥雄蕊和根尖细胞的超微结构的变化;最后,针对神经生物学和脑科学的微观图谱构建,使用基于自动条带收集超薄切片机的高速电镜三维系统,重建了大鼠睾丸精细胞和斑马鱼幼鱼全脑微观连接组,为探索神经系统精细的功能调控提供了结构性的证据。论文的主要成果和贡献如下:

1. 面向植物样品的三维光学显微数据的快速重建算法及分析应用。针对光学显微镜拍摄图像在三维重建过程中遇到的光学畸变和像差等问题,本文基于深度学习设计了一套三维光学显微数据中超微结构的快速重建算法。首先将得到的三维序列光镜图像进行去噪、增强及反卷积处理后,通过改进基于transformer模型的识别网络,对不同类型的组织结构进行了重建。本研究将算法应用于使用结构光照明显微镜成像的不同生长阶段的拟南芥幼苗叶片细胞的内质网时空数据和光片荧光显微镜成像的不同品系杨树种子和油松茎尖分生组织三维数据,定量分析了不同生长阶段内结构的差异,统计和验证了内质网对环境胁迫的动态响应以及为植物组织内多细胞类型的精确空间排列和拓扑结构提供了宝贵的见解,可作为植物研究的数字化参考。

2. 面向植物样品的三维电子显微数据的密集重建算法及分析应用。本文提出了一套基于体电子显微镜序列图像的密集重建算法,对组织细胞内超过10种不同类型的细胞器进行了有效的识别和区分。该框架确保结果的端到端直接输出,包括拓扑网络和形态学分析。本研究将算法应用于使用连续切片自动收集系统和扫描电子显微镜成像的不同发育时期的拟南芥雄蕊数据和拟南芥根尖数据,重建结果的分析证明了细胞的多样性以及具有相似特征的细胞倾向于在特定区域聚集。同时为细胞发育提供了全面的3D视图,这将有助于理解植物细胞生长和分化背后的分子机制。

3. 超大体量三维电镜数据的分布式计算框架设计及斑马鱼全脑连接网络分析。针对大体量动物组织的重建面临的海量数据处理等问题,本文设计了一套应用于超大体量数据的大规模神经环路算法重建的计算框架。
本研究工作采用分治策略,对大体量数据进行分块计算以降低算力需求,同时通过搭建分布式计算框架,提升重建速度。通过设计的体块合并算法,将分布式计算得到的子块结果通过合并算法恢复原始体块体积,从而完成超大大体量数据的高效率重建。同时,利用斑马鱼神经组织的结构特征,研究序列切片电镜图像的切片内相似性和切片间差异性,建立符合神经结构形态的计算模型。经过进一步人工校验,实现斑马鱼幼鱼全脑微观连接组的绘制工作。本研究将算法应用于与中国科学院上海神经所合作获得的国际首套带生物学标记的脊椎模式动物斑马鱼幼鱼的具有突触分辨率的全脑电镜数据上。目前基本完成斑马鱼全脑电镜数据的密集重建工作;完成了全脑17万多细胞核自动分割与人工校验,发现细胞核形貌具有高度多样性;单侧视网膜细胞核自动分割;全脑与单侧视网膜突触PSD、囊泡带、线粒体自动分割;完成了斑马鱼全脑神经元的算法重建;完成了超百个蓝斑去甲肾上腺素能神经元的突触前神经元胞体追踪,发现单个蓝斑神经元即接受广泛输入;不仅蓝斑神经元存在共同输入,而且其与另外两种单胺神经元(多巴胺能与五羟色胺能)接受共同输入支配,即存在若干神经元同时与三种单胺神经元形成突触连接。

Other Abstract

Reconstructing biological tissues at the microscopic scale is crucial for gaining a deep understanding of tissue structure, function, and disease. Traditional two-dimensional imaging analysis has limitations in capturing the complex three-dimensional structure of tissues, which is essential for elucidating the role of tissues in health and disease. In recent years, advances in microscopic tissue reconstruction techniques, such as confocal microscopy, two-photon microscopy, electron microscopy, and tissue clearing methods, have enabled researchers to reconstruct and analyze tissues in three dimensions with unprecedented resolution and depth. These techniques provide new avenues for studying tissue development, morphogenesis, and disease mechanisms, as well as facilitating drug discovery and testing.

This paper focuses on the research of automatic reconstruction and analysis methods for ultra-microstructures of biological tissues at the intersection of cell biology, plant biology, neurobiology, and brain science, based on microscopic imaging techniques and large-scale data processing techniques using artificial intelligence. Firstly, to address the differences in three-dimensional structures of cells from different strains in cell biology, this study investigated the structural differences of the endoplasmic reticulum in Arabidopsis thaliana leaves at different growth stages and the spatial distribution and three-dimensional morphology of seed and stem apical meristem tissues of different poplar strains using structured illumination microscopy and light sheet fluorescence microscopy. Secondly, to explore the relationship between cellular and subcellular structures and functions in plant biology, this study used an automated tape-collecting ultramicrotome system and scanning electron microscopy to investigate the changes in ultrastructure of Arabidopsis anthers and root tip cells at different developmental stages and locations. Finally, to construct microscopic atlases in neurobiology and brain science, this study used a high-speed electron microscopy 3D system based on an automated tape-collecting ultramicrotome to reconstruct the microscopic connectome of rat testicular spermatogenic cells and the whole brain of zebrafish larvae, providing structural evidence for exploring the precise functional regulation of the nervous system. The main achievements and contributions of this paper are as follows:

1. Fast Reconstruction Algorithm and Analytical Application for 3D Optical Microscopy Data of Plant Samples. To address the issues of optical distortion and aberration encountered during the 3D reconstruction process of optical microscope images, this paper designs a set of fast reconstruction algorithms for ultra-microstructures in 3D optical microscopy data based on deep learning. First, the obtained 3D sequence of optical microscope images undergoes denoising, enhancement, and deconvolution processing. Then, different types of tissue structures are reconstructed through an improved recognition network based on the transformer model. this study applied the algorithm to the spatiotemporal data of the endoplasmic reticulum in Arabidopsis thaliana seedling leaf cells at different growth stages imaged by structured illumination microscopy, and the 3D data of seed and stem apical meristem tissues of different poplar strains imaged by light sheet fluorescence microscopy. This study quantitatively analyzed the differences in structures at different growth stages, statistically validated the dynamic response of the endoplasmic reticulum to environmental stresses, and provided valuable insights into the precise spatial arrangement and topological structure of multiple cell types within plant tissues. This can serve as a digital reference for plant research.

2. Dense Reconstruction Algorithm and Analytical Application for 3D Electron Microscopy Data of Plant Samples. This paper proposes a set of dense reconstruction algorithms based on serial block-face scanning electron microscopy images, which effectively identifies and distinguishes more than 10 different types of organelles within tissue cells. The framework ensures end-to-end direct output of results, including topological networks and morphological analysis. This study applied the algorithm to data from Arabidopsis thaliana anthers at different developmental stages and Arabidopsis root tips, which were imaged using an automated tape-collecting ultramicrotome system and scanning electron microscopy. The analysis of the reconstruction results demonstrated the diversity of cells and the tendency of cells with similar characteristics to cluster in specific regions. At the same time, it provided a comprehensive 3D view of cell development, which will help understand the molecular mechanisms behind plant cell growth and differentiation.


3. Design of a Distributed Computing Framework for Ultra-Large-Scale 3D Electron Microscopy Data and Analysis of the Whole-Brain Connectome in Zebrafish.  To address the challenges of processing massive data in large-scale animal tissue reconstructions, this paper designs a computational framework tailored for the reconstruction of large-scale neural circuits from ultra-large datasets. This study employs a divide-and-conquer strategy, segmenting large datasets into smaller blocks to reduce computational demand. Additionally, by establishing a distributed computing framework, the reconstruction speed is significantly enhanced. A designed block-merging algorithm is used to combine the sub-block results obtained from distributed computing, restoring the original block volumes, thereby achieving efficient reconstruction of ultra-large-scale data.
Furthermore, leveraging the structural characteristics of zebrafish neural tissue, the study examines the intra-slice similarity and inter-slice variability of serial electron microscopy images, establishing a computational model that conforms to the morphology of neural structures. After further manual verification, the comprehensive mapping of the zebrafish larval whole-brain microconnectome is accomplished.
The algorithms were applied to the world's first whole-brain electron microscopy dataset of a vertebrate model organism, the zebrafish larva, with synaptic resolution, obtained in collaboration with the Shanghai Institute of Neuroscience, Chinese Academy of Sciences. The dense reconstruction of the zebrafish whole-brain electron microscopy data has been largely completed, including the automatic segmentation and manual verification of over 170,000 cell nuclei, revealing a high diversity in nuclear morphology. This also includes the automatic segmentation of unilateral retinal cell nuclei; the automatic segmentation of whole-brain and unilateral retinal synaptic PSDs, vesicle zones, and mitochondria; the algorithmic reconstruction of zebrafish whole-brain neurons; and the tracing of the presynaptic neuronal somata of over a hundred locus coeruleus noradrenergic neurons, which revealed extensive inputs to individual locus coeruleus neurons. It was found that not only do locus coeruleus neurons share common inputs, but they, along with two other types of monoaminergic neurons (dopaminergic and serotonergic), receive common input regulation, indicating that several neurons simultaneously form synaptic connections with these three types of monoaminergic neurons.

Keyword光学显微镜,体电子显微镜,多模态成像,三维重建,自动分割算法
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/57219
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
刘家正. 面向生物多模态显微成像数据的大规模三维重建方法及应用[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
刘家正-博士学位论文.pdf(72164KB)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
[刘家正]'s Articles
Baidu academic
Similar articles in Baidu academic
[刘家正]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[刘家正]'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.