CASIA OpenIR  > 类脑智能研究中心
面向神经突触连接组的深度学习算法研究及应用
刘静
Subtype博士
Thesis Advisor韩华
2021-11-25
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Discipline模式识别与智能系统
Keyword神经突触 连接组 电子显微镜 深度学习 重建
Abstract

神经突触连接组的绘制是理解大脑工作原理,探究脑疾病发生机制的重要手段。随着序列扫描电子显微镜成像技术的快速发展,解析神经元间的三维突触连接,绘制纳米尺度的大脑线路图已成为可能。然而,不同的脑区组织以及成像方式均有可能导致特异性的数据特点和表现形式,直接应用相同的深度学习算法难以取得满意的效果。

基于扫描电子显微镜序列图像,本文首先基于卷积神经网络构建了神经突触连接组的算法基础,包括突触、线粒体的自动重建算法以及神经元的自动追踪算法。随后,面向小鼠听觉系统不同功能环节的具体科学问题,针对不同脑区、不同成像方式的数据设计了定制化的识别算法,以满足不同场景任务下的重建需求。最后,基于上述算法分别探究了小鼠耳蜗带状突触和听觉皮层多位点突触的组织结构和连接模式,为神经突触实现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. 提出了一种基于卷积神经网络的两步式突触、线粒体识别算法。
为了解决现有突触识别算法不适用强各向异性数据的问题,本文提出了一种新型的两步式突触识别算法:首先应用 Mask R-CNN 在序列图像上检测和分割突触,之后利用突触在相邻层间的连续性设计了 3D connection 算法,实现了强各向异性数据中突触的快速和准确重建。在 Syn-SEM 数据集上验证算法的有效性,本方法的检测性能(精确率: 73%,召回率: 72%)均优于其他方法。针对现有方法对狭长线粒体检测不完整的问题,本文基于两阶段式的目标检测网络设计了一种迭代分割子网络,通过自动的迭代分割检测准确的线粒体膜边界。在 Mito-SEM 数据集上,本方法的分割性能(精确率: 94%,召回率: 84%)超过了目前大部分算法。最后,应用上述算法重建了小鼠皮层 40 × 40 × 50 um3中所有的突触和线粒体,并在结构上统计和验证了突触在三维空间的随机分布,以及线粒体间精细的纳米管道结构的存在。

2. 提出了一种基于 convLSTM 的神经元追踪算法。
针对现有神经元追踪算法自动化程度不高或不适用于强各向异性数据的问题,本文结合 2D 卷积神经网络和 convLSTM 设计了一种神经元追踪网络,同时学习层内和层间的特征,避免了 3D 卷积操作带来的特征空间和图像空间尺度不一致的问题。在训练过程中,引入一种新型的迭代训练模式,通过模拟真实场景的数据,一定程度上降低了追踪过程中的误差累计。针对大体量数据中局部配准精度不高的问题,设计了一个基于深度特征的偏移估计和校正模型,实现了追踪过程中的实时局部配准。该方法在 SEM 数据集和 CREMI 数据集上的 ARE 和VI 指标均超过了现有的神经元追踪方法,证明了该算法的有效性。

3. 面向小鼠耳蜗带状突触,建立了一种基于 3D Detection Network 的突触识别算法,量化分析结果发现了异质性突触功能精细调节的结构性证据。
耳蜗作为听觉系统的感受器,可以将声音信号转换成电信号,而内毛细胞就是耳蜗中负责声音感知的一类细胞。然而,目前内毛细胞中带状突触和其他细胞器间的结构联系尚不清楚。本文采用序列断面扫描电子显微镜这一成像技术来探究小鼠内毛细胞的亚细胞结构,并结合成像特点应用 3D 卷积神经网络设计了多类超微结构的自动识别算法。针对带状突触特殊的生物学结构特征,本文创新性地提出一种 3D Detection Network 实现了带状突触的自动识别和重建,识别精确率和召回率分别为 92% 和 88%。对于内毛细胞中线粒体的重建,本文设计了一种基于 3D U-Net 和 3D connection 算法的自动重建流程,分割精确率和召回率均达到 97%。将上述算法应用到两只小鼠耳蜗的电镜数据中,共获得了 34 个完整的内毛细胞,并对两排内毛细胞交错排列的方式进行了定量化的描述。对带状突触和线粒体进行全细胞范围内的结构量化,结果显示带状突触梯度和线粒体组织模式在两排细胞间存在显著差异,相关性分析揭示了带状突触和线粒体间“位置特异”的相关性关系。

4. 面向小鼠听觉皮层多位点突触,建立了一种多位点突触快速定位和分类的算法,量化分析结果发现一种特殊的突触连接模式与恐惧记忆相关。
听觉皮层作为听觉系统的信息处理中心,突触连接模式的变化为听觉记忆的编码和存储提供了结构基础。已有的研究表明新形成的突触可能与旧突触共存构成“一对多”或“多对一”的多位点突触,然而这种特殊的突触连接模式与记忆到底有何种关系并不清楚。本文以小鼠听觉恐惧学习为实验范式设置实验组和对照组,采用序列切片电子显微镜成像技术获得了 3 组小鼠听觉皮层共 2.8× 105um3数据,应用两步式突触识别算法自动重建了超 160,000 个突触,并结合多种超微结构特征设计了多位点突触的定位和分类算法。统计结果显示听觉恐惧学习没有改变突触的分布密度,但“一对多”的多位点突触比例显著增多。神经元密集重建结果显示这种“一对多”的多位点突触有 85% 以上都连接到了不同树突主干上。建模结果显示这种多树突的突触连接模式大大提高了信息存储容量,可能代表了一种新的突触记忆模式。


 

Other Abstract

Mapping synaptic connectome is an important technology of understanding the principle of brain as well as exploring the mechanism of brain diseases. With the rapid development of serial section scanning electron microscopy technology, analyzing the 3D synaptic connections between neurons and mapping the nanoscale brain circuit have become possible. However, due to the distinct image characteristics caused by different brain tissues and imaging methods, it is difficult to achieve satisfactory performances by applying the constant deep learning algorithms.

Based on serial section scanning electron microscopy images, this paper firstly developed automatic synapse, mitochondrion reconstruction algorithms and neuron tracing algorithm based on convolutional neural networks. Next, be geared to the needs of specific scientific problems in mouse auditory system, a set of customized algorithms were designed for images from different brain regions and different imaging methods in order to meet the demand of different reconstruction tasks. Finally, we applied above algorithms and studied the organizations and configurations of ribbon synapses in mouse cochlear and multi-contact synapses in auditory cortex, providing the structural evidences of a fine-tuned synaptic function. The main achievements and contributions of the paper are as follows:

1. Automatic detection algorithms for synapses and mitochondria based on convolutional neural network were proposed.

In order to solve the problem that the existing synapse detection algorithms were not suitable for highly anisotropic data, this paper proposed a novel two-step synapse detection method. Firstly, we utlized Mask R-CNN to obtain the segmentation and detection results of synapses on 2D slices. Then, a 3D connection algorithm was designed based on the continuity information of synapses in adjacent layers, which realized the fast and accurate synapse detection in highly anisotropic data. The experimental results on Syn-SEM dataset demonstrate that the proposed method achieve superior detection performances (Precision: 73%, Recall: 72%) than other methods. To tackle the problem of incomplete detection of long and narrow mitochondria by existing methods, an recursive segmentation subnetwork was added to the two-stage instance segmentation network to obtain more accurate results by iteratively refining the mitochondrial membrane. The segmentation performance (Precision: 94%, Recall: 84%) on Mito-SEM dataset suggest that the porposed method achieve state-of-the-art results. We reconstructed synapses and mitochondria based on above methods in 40  × 40 × 50 um3 volume from mouse cortex, and we made statistical analysis to validate the random spatial distribution of synapses and the fine nanotunneling structure of mitochondria.

2. A neuron tracing algorithm based on convLSTM was proposed.

To tackle the problems of existing neuron tracing algorithm, such as low level of automation and unsuitability for highly anisotropic data, this paper proposed a neuron tracing network by combining 2D CNN and convLSTM, which can explore the intra-slice and inter-slice features simultaneously and avoid the mismatches on the object scale between feature and image spaces. A novel recursive training method was introduced into the training process, which can reduce the impact of error accumulation by simulating the data from real scenes. To reduce the effects of local misalignments in large-scale data, a deep features-based shift estimation and correction module was designed to calibrate linear deformation during neuron tracing process. ARE and VI metrics of this approach exceed existing neuron tracing methods on SEM and CREMI datasets, which demonstrates the effectiveness.

3. A ribbon synapse detection algorithm based on 3D Detection Network was proposed, and structural evidence of a fine-tuned cytoarchitecture for heterogeneous synaptic function in cochlear inner hair cell was investigated.

As the receptor of auditory system, cochlea can convert sound signals into electrical signals. Inner hair cells (IHCs) are responsible for sound perception. However, the structural relationship between ribbon synapses and other organelles in IHCs is still unclear. In order to explore the subcellular structure of IHCs, this paper used the serial block-face scanning electron microscope for imaging and designed a set of automatic IHC ultrastructures detection algorithm based on 3D convolutional neural network. According to the biological features of ribbon synapses, a novel 3D instance segmentation network was proposed to obtain the accurate detection of ribbon synapses, and yielded 92% and 88% detection precision and recall.  An automatic reconstruction algorithm for mitochondria was developed based on 3D U-Net and 3D connection algorithm, and yielded 97% segmentation precision and recall. A total of 34 IHCs were reconstructed from electron microscopy volume of two mice. And morphological quantification of staggered arrangement of IHCs was carried out. Cell-wide structural quantification of ribbon synapses and mitochondria suggests significant differences of ribbon morphology gradient and mitochondrial organization between adjacent IHCs in staggered pairs. Correlation analysis argues for a "location-specific" correlation between ribbon synapses and mitochondria.

4. An automatic location and classification method of multi-contact synapses was proposed, and a special pattern of synaptic connections was found to be related with the fear memory.

As the information processing center of auditory system, the change of synaptic connection patterns in auditory cortex provides the basis for the coding and storage of memory. Existing methods demonstrate that the newly formed synapses may coexist with the old synapses to form "one-to-many" or "many-to-one" multi-contact synapses. However, the relationship between the synaptic connection pattern and fear memory is unclear. This paper set up the conditioned groups and the control groups with the auditory fear conditioning as the behavioral paradigm. 2.8 ×105 um3 electron microscopy data from mice of 3 group pairs were yielded by using the serial section electron microscope imaging technology. Upon reconstructing over 160,000 synapses, this paper designed a location and classification method of multi-contact synapses based on the features of a variety of ultrastructures. The statistical analysis shows that fear learning don't change the density of synapses but significantly increases the proportion of "one-to-many" multi-contact synapses. Dense neuron reconstructions show that more than 85% of these synapses are connected to different dendrites. Mathematical modeling results suggest that the synaptic connection pattern of the multi-dendrites greatly improves the information storage capacity, which may represent a new synaptic memory engram.

Pages154
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/46604
Collection类脑智能研究中心
Recommended Citation
GB/T 7714
刘静. 面向神经突触连接组的深度学习算法研究及应用[D]. 中国科学院自动化研究所. 中国科学院大学,2021.
Files in This Item:
File Name/Size DocType Version Access License
面向神经突触连接组的深度学习算法研究及应(27979KB)学位论文 开放获取CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[刘静]'s Articles
Baidu academic
Similar articles in Baidu academic
[刘静]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[刘静]'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.