CASIA OpenIR  > 毕业生  > 博士学位论文
基于序列切片电子断层成像的神经元细胞器三维重建算法研究
常胜
2023-05
Pages134
Subtype博士
Abstract

神经元细胞器三维结构的重建作为微观脑图谱绘制中至关重要的一环,对于解析神经元结构,研究神经元之间信息传递方式,探索神经功能以及新一代人工智能有着非常重要的意义。然而目前主流的序列切片电子显微镜成像技术分辨率有限,且重建的体块存在着明显的各向异性;电子断层成像技术重建体量有限,无法获得完整的神经元细胞器三维结构。因此对结合了序列切片电子显微镜成像技术大重建体量特点和电子断层成像技术高分辨率特点的序列切片电子断层成像技术展开研究,实现在亚纳米级分辨率下重建微米级体量的完整神经元细胞器三维结构是非常有必要的。

本文旨在基于序列切片电子断层成像技术,设计一套从原始电镜图像处理,到厚切片数据对齐,再到目标结构分割的完整的神经元细胞器三维重构方法。首先,针对低电子剂量拍摄造成的电子显微镜图像中的噪声,研究了它们产生的原因和去除方法;其次,针对厚切片自身的形态变化,以及切片、成像、断层扫描重建过程中存在的各种损失,研究了一套包含序列体块对齐、块间损失修补在内的完整的序列体块电子断层成像三维重建工作流程,并利用该流程重建分析了野生型大鼠大脑的前额叶皮层中的化学突触;最后结合多种生物电镜成像技术,重建了年轻和年老大鼠视网膜中带状突触的三维结构,并研究了衰老对大鼠视网膜中带状突触的影响。论文的主要成果和贡献如下:

1. 针对电子显微镜图像中不可避免的信号依赖噪声,本文提出了一种基于方差稳定化原理的去噪网络。首先根据电子显微镜图像中噪声产生的原因,引入了与之对应的噪声模型,并利用标准样品验证了噪声模型的有效性。然后基于方差稳定化原理提出了相应的去噪网络,以实现电子显微镜图像中噪声的去除。去噪过程主要包含两个阶段,分别是方差稳定化阶段和去噪阶段。在方差稳定化阶段利用深度神经网络代替传统方差稳定化变换中的泰勒公式进行近似计算,从而在实现信号与噪声分离的同时保持了原始图像域中的信号,减少了忽略泰勒公式中余项带来的近似误差,也避免了正反方差稳定化过程中的误差累积。在去噪阶段利用带有通道注意力模块的U-Net网络来实现对图像中内容特征的表征,从而保留了更多的图像细节信息,在保证去噪效果的同时增强了电子显微镜图像中的超微结构。与现有的其他电子显微镜图像去噪方法相比,本文提出的方法在客观评估和视觉效果上都更具有优势。

2. 针对序列厚切片间较大的形态变化和块间信息损失,本文提出了一套完整的序列切片电子断层成像三维重建算法。对于形貌具有明显差异的序列电子断层重建体块的对齐,本文使用多尺度图像特征进行匹配,并利用具有置信度传播的弹性变形算法(BP- Elastic)从粗到细地对齐它们。为了恢复电子断层重建体块中损失的信息,本文根据对齐后相邻体块之间的像素变化来估计损失的图像数量,再利用预训练插值网络,基于蒸馏学习的方法,设计了一个适用于小样本电子断层成像数据的厚切片片间缺失信息生成网络,从而实现了在z方向更具有连续性的完整体块的重建。最后基于提出的工作流程,本文完成了多组野生型大鼠大脑的前额叶皮层中完整突触的亚纳米级尺度三维重建,并统计分析了突触内囊泡与突触前膜之间的空间位置关系,为其他亚细胞结构的重建分析提供了有效手段。

3. 基于多种生物电镜成像技术,本文重建了年轻和年老的大鼠视网膜中的带状突触,研究了衰老对大鼠视网膜中带状突触的影响。实验结果表明随着年龄的增长,大鼠视网膜中带状突触的数量减少,而具有自噬结构(电镜图像中为黑色斑块)的异常带状突触以及自噬线粒体的数目增加,但是单个带状突触的表面积、带状突触中条带的体积以及带状突触中线粒体的数目和体积并没有显著的差异,只是带状突触中黑色斑块的数量明显上升。这说明年轻和年老大鼠视网膜中带状突触的基本结构是相同的,只是衰老会使大鼠视网膜中的带状突触数量减少,且更易出现病变的带状突触,从而可能导致老年视觉障碍及退行性疾病。实验结果有望为研究视网膜衰老及相关疾病提供技术手段。

Other Abstract

The reconstruction of three-dimensional (3D) structure of neuronal organelles is a pivotal aspect in the micro-scale mapping of the brain, with profound implications for analyzing neuronal architecture, elucidating information transmission among neurons, investigating neural functionalities, and advancing the field of next-generation artificial intelligence. However, current mainstream imaging techniques, such as serial section electron microscopy, have inherent limitations in resolution and exhibit noticeable anisotropy in the reconstructed volumes. Likewise, electron tomography suffers from limited imaging volume, rendering it insufficient for complete 3D reconstruction of neuronal organelles. Hence, it is imperative to explore serial electron tomography techniques that can achieve sub-nanometer resolution and micrometer-scale volume, facilitating comprehensive reconstruction of the 3D structure of neuronal organelles.

The objective of this article is to devise a comprehensive methodology for reconstructing the 3D structure of neuronal organelles utilizing serial section electron tomography. The proposed approach encompasses the entire process, starting from the processing of raw electron microscopy images, followed by the alignment of thick sections, and culminating in the reconstruction of target structures. Initially, the issue of noise in electron microscope images, induced by low electron dose, is thoroughly examined, including its underlying causes and corresponding removal techniques. Subsequently, a complete workflow for serial section electron tomography 3D reconstruction is formulated, incorporating serial section alignment and inter-section loss correction strategies to account for morphological changes in thick sections and various losses incurred during sectioning, imaging, and tomographic reconstruction. Employing this workflow, the reconstruction and analysis of conventional chemical synapses are conducted. Lastly, by integrating electron microscopy, sequential section electron microscopy, and sequential section electron tomography techniques, the 3D structure of ribbon synapses in the retinas of young and old rats is reconstructed, enabling the investigation of the impact of aging on ribbon synapses in rat retinas and the exploration of potential causes of visual impairment and degenerative diseases in the elderly. The main achievements and contributions of my paper are summarized as follows:

1. To mitigate the inevitable signal-dependent noise in electron microscopy (EM) images, this paper delves into the underlying causes of noise and proposes targeted denoising strategies. Firstly, a corresponding noise model is introduced, grounded on the cause of noise in EM images, and its effectiveness is verified using standard samples. Subsequently, based on the principle of variance stabilization, a corresponding denoising network is proposed to effectively remove the noise in EM images. The denoising process entails two main stages: variance stabilization and denoising. In the variance stabilization stage, a deep neural network is utilized to replace the traditional Taylor formula for approximate calculations, facilitating the separation of signals and noise, while maintaining the signal in the image domain. This approach also reduces the approximation error caused by the remainder term in the Taylor formula, thus avoiding error accumulation in the forward and inverse variance stabilization processes. In the denoising stage, a U-Net network with channel attention modules is employed to capture the content features of the image, preserving crucial image details and enhancing the ultrastructure in EM images, while ensuring effective denoising. In comparison to existing denoising methods for EM images, the proposed approach in this paper demonstrates advantages in objective evaluation and visual effect.

2. To address the challenges posed by significant morphological changes and loss of inter-section information in serial-section electron tomography, this paper presents an intelligent method for 3D reconstruction in this context. For aligning tomographic sections with notable morphological differences, the proposed method leverages multi-scale features from multi-resolution images and employs an elastic deformation algorithm coupled with belief propagation to achieve coarse-to-fine alignment. To restore missing information in the reconstructed tomographic sections, the paper estimates the number of lost images based on pixel changes between adjacent sections after alignment. Subsequently, a thick-section inter-missing information generation network is designed, utilizing pre-trained interpolation networks and distillation learning methods, to effectively fill the gaps and achieve greater continuity in the z-direction. Finally, utilizing the aforementioned methodology, this study successfully achieves sub-nanometer-scale 3D reconstruction of multiple complete synapses in the fourth layer of the barrel cortex of a wild-type rat and conducts statistical analysis on the spatial relationship between synaptic vesicles and presynaptic membranes. The proposed workflow and method for 3D reconstruction and analysis of serial-section electron tomography provides an effective approach not only for studying synapses in the barrel cortex of a wild-type rat, but also for reconstructing and analyzing other subcellular structures.

3. By employing advanced three-dimensional reconstruction techniques based on diverse biological electron microscopy imaging methods, this study aims to investigate the impact of aging on ribbon synapses in the retina of rats. Through the use of sophisticated image stitching, registration, and segmentation algorithms, this paper has successfully reconstructed ribbon synapses in both young and elderly rats. The experimental findings reveal a notable decrease in the number of ribbon synapses in the retina of rats as they age, accompanied by an increase in the number of abnormal ribbon synapses exhibiting autophagic structures and autophagic mitochondria. However, there are no significant differences in the basic structural parameters of individual ribbon synapses, such as surface area, volume of bands, and number and volume of mitochondria, except for the increased number of autophagic structures observed in ribbon synapses with aging. These results suggest that aging can impact both the quantity and quality of ribbon synapses in the retina of rats, which may potentially lead to visual impairment and degenerative diseases in the elderly. Overall, the findings of this study provide valuable insights and technical approaches for the study of retinal aging and associated diseases.

Keyword电子显微镜 序列电子断层扫描 去噪 拼接配准 损失信息生成
Language中文
Sub direction classification医学影像处理与分析
planning direction of the national heavy laboratoryAI For Science
Paper associated data
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/52218
Collection毕业生_博士学位论文
脑图谱与类脑智能实验室_微观重建与智能分析
Recommended Citation
GB/T 7714
常胜. 基于序列切片电子断层成像的神经元细胞器三维重建算法研究[D],2023.
Files in This Item:
File Name/Size DocType Version Access License
基于序列切片电子断层成像的神经元细胞器三(35861KB)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
[常胜]'s Articles
Baidu academic
Similar articles in Baidu academic
[常胜]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[常胜]'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.