毕业生知识库

脑细胞之间的联接关系是大脑实现认知功能的生理基础。重建微观尺度下的脑组织联接关系是探究脑运行机理的重要手段之一。纳米尺度下的脑微观结构重建技术能够有效地恢复神经组织的微观三维结构，为后续的神经科学和生物学研究奠定坚实的基础。脑微观结构重建环节中的序列切片电镜图像配准过程是限制脑重建工程规模的技术瓶颈之一。开展脑序列切片图像配准精度理论研究对提高配准算法的配准精度，改善序列切片图像配准质量，获得可靠的脑微观结构三维重建结果至关重要。

本文是脑微观结构重建研究工作的一部分，探究可合理应用于神经组织序列切片电镜图像配准过程的配准精度分析模型和基于精度分析结论改进的图像配准算法。通过建立的配准精度分析模型，本研究分析了在脑组织序列切片电镜图像配准流程中影响配准精度的关键因素，针对性地改进了图像配准算法，获得了质量更高的图像配准结果。

本文的研究分为两步进行：首先，针对在配准过程中需要求解图像局部间的平移向量问题，本文提出了一种结合了神经元结构形态的配准精度分析模型。该模型综合考虑了神经元结构和切片过程对配准精度造成的影响。基于该模型，本文从数学上分析了局部图像块间的配准精度和切片厚度、结构大小以及结构形状之间的关系。实验结果表明，根据模型推导得到的配准精度关系适用于真实的神经组织序列切片电镜图像配准方法。其次，本文结合上述推导得出的配准精度分析结果改进了已有的图像配准算法，并成功地将其应用于真实脑组织序列切片图像配准工程中。改进的配准方法分别考虑了神经结构的尺寸信息和形状信息，并设计了额外的结构信息处理步骤。在测试数据集上的配准结果显示，相比较已有的配准方法，本文提出的改进配准方法可以在真实神经组织序列切片图像数据上取得更好的图像配准效果。

The connectivity between neurons is the physiological basis for cognitive functions. Reconstruction of nerve tissue connections at microscale is an essential tool for investigating the mechanism of brain operation. Nanoscale reconstruction techniques can effectively restore the three-dimensional microscale structure of neural tissue, laying a solid foundation for subsequent neuroscience and biological research. One of the technical bottlenecks limiting the scale of brain reconstruction projects is registering serial section electron microscopy images in the reconstruction process. Theoretical studies on the registration accuracy of serial section images of nerve tissue are essential to improve the accuracy of registration algorithms, improve the quality of serial slice image registration results, and obtain reliable 3D reconstruction results of neuron structures.

This paper is part of the research on brain microscale reconstruction to investigate the registration accuracy analysis model and improved image registration algorithms based on the derived accuracy conclusion, which can be reasonably applied to the electron microscopy image registration process of neural tissue serial sections. By using the accuracy analysis model developed, this study analyses the key factors affecting the accuracy of registration in the microscale reconstructure of nerve tissue and optimize the registration algorithm to obtain better registration results.

The study is divided into two steps: first, to estimate the local translation vectors' distribution in the registration process, an registration accuracy analysis model incorporating the structural morphology of neurons is proposed. The model considers the influence of neuron structure and the slicing process on the registration accuracy. Based on this model, this paper mathematically analyse the relationship between the registration accuracy and the slice thickness, structure size, and structure shape between local image blocks. The experimental results show that the registration accuracy relationship derived from the model is applicable to the registration methods for electron microscopy images of neural tissue sequence sections. Secondly, this paper improves the existing image registration algorithm by combining the above-derived registration accuracy analysis results and successfully applies it to a real serial section image registration process. The improved registration method considers the size and shape information of the neural structures respectively, and devises additional steps to process the structural information. The registration results on the test dataset show that the improved registration method can achieve better image registration results on actual serial section electron microscopy image data compared to the existing methods.