CASIA OpenIR
生物组织序列切片图像配准技术研究
舒畅
Subtype博士
Thesis Advisor韩华 ; 陈曦
2020-06-01
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计算机应用技术
Keyword医学图像配准 序列图像配准 神经元的三维重建 序列切片图像
Abstract

虽然科学技术在不断进步,但是大脑的运作方式仍然是未解之谜。结构决定功能,了解大脑的运作方式需要在纳米级的分辨率下对神经元成像。现有的三维成像设备容量有限,无法容纳大多数的生物样品。因此,对生物组织进行切片和成像,再从得到的图像序列中重建出神经元的三维结构,是获取大体量神经环路三维影像数据的有效手段。

然而在切片过程中,生物组织结构的连续性会被破坏,相邻的切片之间会引入扭曲、褶皱、伸缩等非线性形变。整个图像序列需要配准,如果不能较好地处理这些形变,将会影响后续的生物学分析。因此,研究能够恢复图像序列沿轴向的连续性的序列图像配准方法至关重要。

本文旨在设计高效的序列图像配准方法,以满足大规模神经元三维重建任务的需要。针对现有方法的不足之处,本文结合具体的生物学特点来优化对应算法的设计,突破大体量数据配准精度低、周期长、自动化程度低等技术瓶颈。

本文的主要工作和创新点归纳如下:

1. 提出了一个基于分治策略的两两图像高速高精度配准方法。针对前人方法在高速和高精度上难以共存的矛盾,本文充分利用生物组织序列切片的形变特点,借助分而治之的思想,对原问题进行了充分简化,据此设计了一个双重网络来实现高速高精度的图像配准,并且通过无监督学习的方式避免了昂贵的人工标注,增加了网络在实际场景中的应用价值。在Cremi数据集上,该方法以0.1秒每帧的速度实现0.758的配准精度(结构相似度指标),在速度和精度上均优于同类算法。进一步的实验证明,该方法能够作为序列图像配准任务的子环节,用以快速提取图像之间的对应关系。相比于序列图像配准中常用的对应点提取方法,该方法能够处理超大分辨率的图像,并且在速度上有数百倍的提升。

2. 提出了一种基于生物结构约束的长图像序列配准解算方法。针对序列图像配准任务中晶圆间存在成像畸变、长序列配准容易发生香蕉效应等问题,本文提出以晶圆为单位进行配准以避免长序列配准导致的误差累积,在晶圆内基于生物结构的连续性施加更为严格的约束以提升配准质量。更为严格的约束条件增加了问题的求解难度。为此,本文针对性地设计了一个非迭代的近似算法来实现快速求解,并且在理论上分析了该算法的最优性条件,在实验上证明了该算法的有效性和高效性。基于此方法,本文以较短的时间实现了大体量(675.8×675.8×181.2立方微米)的斑马鱼数据的配准,恢复出了平滑连续的生物构,为后续的生物学分析打下了坚实的基础。并且通过统计分析,本文指出了未来序列图像配准任务效率提升的潜在方向。

3. 提出了一种适用于大体量多数据类型的序列图像配准方法。现有序列切片图像配准方法主要针对薄或者超薄切片设计,适用面窄,不利于在光电关联等任务上的应用。为此,本文指出切片厚度不均、片间差异大和大数据量是阻碍序列图像配准技术拓展的主要障碍,并提出了切片表面自动修平、稠密的对应点提取、非迭代的全局优化等技术来消除这些障碍,得到了适用于大体量多数据类型的序列图像配准算法。本文通过全面的实验证明了该方法的有效性,并且通过统计分析指出,所施加约束的严格程度是影响序列图像配准算法适用性的主要因素,为后续序列图像配准算法的设计提供了思路。以该算法为基础,本文完成了高精度的大体量生物组织的配准,包括3.9×4.6×5.6立方毫米体量的小鼠胚胎脑数据和13.7×9.0×12.1立方毫米体量的小鼠透明脑数据,恢复出了平滑连续的生物组织结构,为后续的生物学分析提供良好的依据。

Other Abstract

Although science and technology are improving, it remains a mystery as to how brains work. It is well known that the function is determined by the structure. To understand the way brains work, we need to image the structure of neurons at nano-meter resolution. The capacity of existing 3D imaging equipment is too small to accommodate most biological samples. Therefore, imaging serial sections of biological tissue and then reconstructing the three-dimensional structures of neurons therein is an effective way to obtain 3D image data of large volume neural circuits.

In the process of sectioning, the continuity of biological tissue is destroyed, and nonlinear deformations such as distortion, folding, shrinkage and expansion are introduced between adjacent sections. The whole image sequence needs to be registered. The artifacts caused by deformation could affect the subsequent analysis seriously if not properly handled. Therefore, researches on the serial image registration methods which can restore the axial continuity of image sequences are of great significance.

The purpose of this paper is to design an efficient serial image registration method to meet the needs of large-scale neuron 3D reconstruction tasks. In view of the shortcomings of the existing methods, this paper utilizes specific biological features to optimize the design of corresponding algorithms, breaking through the technical bottlenecks like low registration accuracy, long reconstruction period and low automatic level in large-scale registration.

The main work and innovation of this paper are summarized as follows:

1. A fast and accurate pairwise image registration method based on divide and conquer strategy is proposed. To tackle the contradiction between high speed and high precision in the previous methods, this paper makes full use of the deformation characteristics of biological tissue serial sections, simplifies the original problem with divide and conquer strategy, and consequently designs a dual network for high-speed and high-precision image registration. Moreover, unsupervised learning is introduced to avoid costly manual annotation and increases the application value of the network in the real scenes. On the Cremi dataset, our method achieves 0.758 registration accuracy (structural similarity metric) with a speed of 10 fps, outperforming baseline methods in terms of speed and accuracy by a large margin. Further experiments show that this method can be used as a sub-link of serial image registration tasks to quickly extract correspondences between images. Compared with correspondence extraction methods widely used in serial image registration, this method can process images with super-high resolution, and is hundreds of times faster.

2. A solution for long image sequence registration with biological structure constraints is proposed. In order to solve the problems of inter-wafer imaging distortion and banana effect in long image sequence registration, this paper proposes to conduct wafer-wise registration to avoid the error accumulation caused by long sequence registration and impose strict constraints within a wafer based on the continuity of biological structure to ensure registration quality. Stricter constraints increase the difficulty of the problem solving. Therefore, this paper designs a non-iterative approximation algorithm to provide a fast solution, and theoretically analyzes the algorithm's optimization conditions and experimentally proves the algorithm's effectiveness and efficiency. Based on this method, we achieve the registration of zebrafish data with a large volume (675.8×675.8×181.2 cubic micrometer) in a short time. Smooth and continuous biological structures are recovered, which lay a solid foundation for subsequent biological analysis. And through statistical analysis, this paper points out the potential direction of serial image registration tasks' efficiency improvement in the future.

3. A serial image registration method for large volume and multiple data types is proposed. Existing serial-section image registration methods are mainly designed for thin or ultra-thin sections, they have limited applicability and are not suitable for applications like correlative light and electron microscopy tasks. For this reason, this paper points out that section thickness unevenness, large inter-section difference and huge data size are main obstacles to the extension of serial image registration methods, and proposes corresponding technologies like automatic volume flattening, dense correspondence extraction and non-iterative optimization to eliminate these barriers, resulting in a serial image registration method suitable for large volume and multiple data types. This paper conducts comprehensive experiments to verify the effectiveness of this method. Through statistical analysis, we point out that the strictness of imposed constraints is the main factor affecting the applicability of a serial image registration algorithm, providing a thinking for the design of subsequent serial image registration methods. Based on this method, we complete reconstructions of large-scale biological tissues, including a mouse embryo brain with a volume of 3.9×4.6×5.6 cubic millimeter and a cleared mouse brain with a volume of 13.7×9.0×12.1 cubic millimeter. Smooth and continuous biological tissue structures are recovered, which provide a good basis for further biological analysis.

Pages142
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/39229
Collection中国科学院自动化研究所
Recommended Citation
GB/T 7714
舒畅. 生物组织序列切片图像配准技术研究[D]. 中国科学院自动化研究所. 中国科学院大学,2020.
Files in This Item:
File Name/Size DocType Version Access License
签名版5.pdf(13897KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[舒畅]'s Articles
Baidu academic
Similar articles in Baidu academic
[舒畅]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[舒畅]'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.