毕业生知识库

微观连接组学旨在绘制具有突触分辨率的神经元“连接线路图”，正逐渐成为理解大脑工作机制的重要方式。基于序列切片电子显微镜成像技术获取大规模突触水平图像数据并对神经结构三维自动重建是进一步解析神经元连接线路图的必经之路。因此，研究准确高效且符合神经结构特性的三维重建算法，是加速电镜数据处理能力，突破连接组发展瓶颈的关键技术。

为了获取连接线路图，需要恢复两个关键结构：神经元结构以及连接神经元的突触结构。然而，样品制备过程中神经组织的无差别染色、切片过程中三维结构的信息缺失、数据拼接及配准中过程中的非线性形变以及神经结构高度复杂且形态各异等诸多难以避免的物理因素，都影响后续神经结构识别的准确率。针对这两类结构，本文充分结合电镜图像特点与神经结构生物先验信息，将神经结构的形态特点、空间信息以及连接信息融入到算法设计中，从模型结构与数据特点两个方面来建立高效的自动分割算法，支撑大规模微观连接组学的研究发展。论文研究内容主要包括三部分，分别是神经元膜结构分割算法、突触精细结构的三维实例分割算法，并在这两个结果基础上构建符合生物合理性的神经元重建框架。论文的主要工作与贡献归纳如下：

1. 针对目前神经元膜分割算法普遍存在边界结果不连续，膜细节分割不精细的缺点，本文基于深度学习和图模型设计了一套神经元膜结构自动分割算法，通过结合神经结构多尺度信息提高神经元膜结构的识别效果。基于网络感受野与膜结构的生物特点，本文提出的膜分割网络结合上下文残差和亚像素卷积，有效克服了模糊边界的二义性问题，减少膜缺失现象，并融合多尺度特征图来避免膜边界与其他超微结构的混淆。针对网络初始分割结果中膜结构不连续现象，本文使用 lifted multicut 算法，通过远程多尺度信息有效优化了神经元膜结构的分割效果。本文提出的膜分割算法在 ISBI 电镜分割公开挑战赛中的结果接近人类专家准确度，在 243 支参赛队伍中排名第 3。 

2. 针对突触接触面形状不规则且边界不清晰的问题，本文基于神经结构空间一致性信息提出了一种用于端到端突触重建的三维实例分割网络，从而弥补目前两段式重建方法的局限性。本文通过三维卷积核提取与生物结构相一致的区域特征，能同时输出突触的检测结果与分割结果。端到端的训练方式促进了网 络的拟合能力，并显著提高了性能。此外，引入语义分割分支和特征融合模块能 辅助网络的特征表达，进一步改进分割性能。为适配大体量重建需求，本文开发 了基于块的重建策略，通过分块及拼接处理满足大规模连接组数据的重建需要 和分割准确率。在两个电镜数据集上的实验证明，本文提出的网络可以有效提取 符合突触生物结构的三维区域特征，并能大幅减少假阳性突触与三维突触的连 接错误，平均精度比基线方法提升了 6.2%。

3. 针对目前神经元聚合算法忽略了神经结构内在相关性的缺点，本文基于突触及线粒体等超微结构的连接性约束设计了具有生物连接信息的神经元三维聚合框架，避免重建结果对局部膜边界线索的过度依赖。基于超微结构连接信息，本文首先设计了连接约束模型，该模型将超微结构实例分割结果编码为符合生物先验的神经元连接性约束项，有效避免了不合理聚合结果。针对提取的连接约束项可能存在潜在的错误，本文继而设计了联合优化模型，该模型基于超微结构语义分割结果构建了神经结构共聚合算法，有效克服了不理想超微结构的影响。针对线粒体和突触的连接先验不同，本文明确表示了神经元与线粒体及神经元与突触的两种连接模式。本文提出的三维聚合算法可以整合多种连接信息，产生更具神经结构合理性的三维分割结果，并在三个公开数据集上验证了结合生物连接信息能提升 3.8% 的神经元分割性能。

Micro-connectomics, which aims to map neuronal ”wiring diagram” with synaptic resolution, is becoming an important way to understand the working mechanisms of the brain. Based on large-scale synaptic-level image data obtained by serial section electron microscopy imaging technology, three-dimensional (3D) automatic reconstruction of neural structures is a fundamental path to further analyze neuronal connectivity maps. Therefore, the research of accurate and efficient 3D reconstruction algorithms which accords with the characteristics of neural structure is the key technology to accelerate the data processing ability of electron microscopes and break through the bottleneck of the development of connectomics.

In order to obtain wiring diagrams of connections, two key neural structures need to be recovered: the neuronal structure as well as the synaptic structure connecting the neurons. However, the accuracy of subsequent neural structure recognition is affected by many unavoidable physical factors, such as the undifferentiated staining of neural tissue during sample preparation, the missing information of 3D structure during sectioning, the nonlinear deformation during data stitching and alignment, and the highly complex and morphologically diverse of neural structure. To tackle these two types of structures, this paper fully combines the characteristics of electron microscopic images and the biological prior information of neural structures, incorporates the morphological characteristics, spatial information and connectivity information of neural structures into the algorithm design, and establishes an efficient automatic segmentation algorithm in terms of both model architecture and data characteristics to support the development of large-scale micro-connectomics research. The research content of this thesis mainly includes three parts, which are neuron membrane structure segmentation algorithm, the 3D instance segmentation algorithm for synaptic fine structure, and the construction of a biologically plausible neuron reconstruction framework based on these two results. The main work and contributions of the thesis are summarized as follows.

1. To address the disadvantages that the current neuron membrane segmentation algorithms generally have discontinuous boundary results and not fine membrane detail segmentation, this paper based on deep learning and graph model to improve the recognition effect of neuron membrane by combining neural structure multi-scale information. Based on the biological characteristics of network receptive field and membrane structure, the proposed membrane segmentation network combines contextual residuals and subpixel convolution to effectively overcome the ambiguity problem of ambiguous boundaries, reduce membrane missing phenomenon, and fuse multi-scale feature maps to avoid the confusion of membrane boundary with other ultrastructures. To address the membrane structure discontinuity in the initial segmentation results of the network, this paper uses the lifted multicut algorithm to effectively optimize the neuron membrane structure segmentation results by using long-range multiscale information. The results of the membrane structure segmentation algorithm proposed in this paper are close to the accuracy of human experts in the ISBI Electron Microscopy Segmentation Public Challenge, and ranked 3rd among 243 participating teams.

2. To address the problem of irregular shape and unclear boundary of synaptic contact surface, this paper proposes a 3D instance segmentation network for end-to-end synapse reconstruction based on the spatial consistency information of neural structure, so as to compensate the limitations of current two-segment reconstruction methods. In this paper, the regional features consistent with the biological structure are extracted by the 3D convolution kernel, which can output the synapse detection results and segmentation results simultaneously. The end-to-end training approach facilitates the fitting ability of the network and significantly improves the performance. In addition, the introduction of semantic segmentation branches and feature fusion modules can improve the segmentation performance further by auxiliary feature representation of the network. To adapt to the large volume reconstruction demand, a block-wise reconstruction strategy is developed in this paper to meet the reconstruction needs and segmentation accuracy of large-scale connectomics data through chunking and splicing processing. Experiments on two electron microscopy datasets demonstrate that the proposed network can effectively extract 3D regional features consistent with synaptic biology and can significantly reduce false-positive synapses and 3D connection errors, with an average accuracy improvement of 6.2% over the baseline method.

3. To address the drawback that the current neuron agglomeration algorithms ignore the inherent relevance of neural structures, this paper designs a 3D neuron agglomeration framework with biological connectivity information based on the connectivity constraints of ultrastructures, such as synapses and mitochondria, to avoid the dependence of reconstruction results on local membrane boundary cues. Based on the ultrastructural connectivity information, this paper first designs a connectivity constraint model, which encodes the ultrastructural instance segmentation results into connectivity constraint terms that conform to the biological prior, effectively avoiding implausible agglomeration results. To address the potential errors of the extracted connectivity constraints, this paper then designs a joint optimization model, which constructs a neural structure co-agglomeration algorithm based on the ultrastructural semantic segmentation results, effectively overcoming the influence of undesirable ultrastructures. For the different connectivity constraints of mitochondria and synapses, this paper explicitly represents two connectivity patterns, namely, neuron-mitochondria, and neuron-synapse, respectively. The proposed 3D agglomeration algorithm can integrate multiple connectivity information to produce 3D segmentation results with more plausible neural structure, and the combined biological connectivity information is verified to improve the neuron segmentation performance by 3.8% on three public datasets.