毕业生知识库

    猕猴作为一种重要的实验动物，它与人类有许多共同的遗传、大脑组织和行为特征。猕猴脑图谱可以识别解剖学和功能上的不同区域，是生物医学和进化研究的基础。对医学影像和生物成像数据进行多模态和跨尺度的联合利用，不仅可以提供介观连接和多组学图谱，还可以促进转化医学、跨物种比较和计算建模的发展，丰富非人类灵长类的合作资源平台。研究猕猴脑图谱多尺度数据的空间配准方法，确定多尺度集成的评价准则，对于制作用于猕猴脑图谱多尺度集成的标准参考空间具有重要意义。对于生物成像与医学影像间配准方法的研究可以建立生物成像和医学影像间对应关系，基于此可以将带有三维空间坐标参考的空间离散数据嵌入到标准空间，以便于后续各种参考标准整合，充分利用不同模态数据的优势，对大脑进行进一步的联合分析。现存的大量配准方法主要针对单模态的神经系统，而对于在灰度、组织和解剖结构上成像具有显著差异的，数据间图像特征差异较大的脑图谱数据，很少有跨模态、跨尺度配准的方法可以直接应用，并且不同种类的脑图谱数据存在尺度和模态的双重不同，如何将猕猴脑图谱生物成像与医学影像间通过图像配准方法进行集成存在一系列挑战。

    针对这一现状，本文系统地研究了猕猴脑图谱生物成像与医学影像数据间的配准集成的方法，并且给出了标准空间集成的选择标准以及集成效果的衡量指标；以本中心采集的R04猕猴大脑成像数据作为主要数据集，将数据集中的相机断面数据作为中间模态以及集成的标准空间，进一步将猕猴大脑的尼式染色数据以及MRI数据集成到该空间。此外，借助公开的Allen Human Brain Atlas数据集对于本文所提出方法进行了验证。本文的主要内容如下：

    1.多模态跨尺度猕猴脑影像质量控制与数据处理。本文首先介绍了R04猕猴大脑成像数据的实验采集过程，通过质量控制对于数据进行筛选，保留可用数据，用于后续实验；对于不可用数据，针对数据出现的质量问题通过寻因分析寻找实验过程中导致数据出现质量问题的原因，协助实验老师改进实验以达到提高之后实验所产生数据集质量的目的。针对尼氏染色数据的预处理流程目前无统一预处理步骤的现状，构建了一套完整的尼式染色数据预处理方法；对于质量控制后的可用数据将其预处理成可以直接读入神经网络和医学开源工具箱的格式。

    2.多模态跨尺度的猕猴脑图谱配准新方法研究。确定了选择标准空间的准则，基于SyN算法，针对脑图谱数据集成的两个主要问题-三维重建和不同尺度数据间的空间映射，以R04猕猴大脑成像数据为例，设计了一整套实现猕猴大脑生物成像数据和医学影像数据的多模态跨尺度的配准方案，并且给出了不同尺度不同模态数据之间的相互映射方法，初步解决尺度模态双重不同的配准问题。提出使用区域解剖标签重叠的DSC与基于L1损失衡量的Id Score作为正反变换一致性的评价指标。

    3.基于深度学习的多模态跨尺度灵长类脑图谱配准方法研究。为提高处理速度以及空间配准精度，改进基于SyN算法的猕猴脑图谱多模态跨尺度配准方案的关键步骤；分别提出了线性配准网络和基于图像迁移与配准交替优化的非线性配准框架；将图像迁移引入跨模态配准，解决跨模态配准损失函数难以设计的问题；将循环一致性引入正反变换一致性问题，解决图像依次经过正反变换后与原图存在差异的问题，以此来减少空间映射的误差。在保证精度的情况下，快速处理大批数据。最终得到一套同时具备速度和精度的多模态跨尺度猕猴脑图谱配准方案。实验结果显示，本文所提出框架下的方法在R04猕猴大脑成像数据集上相比于SyN算法和VoxelMorph-Diff模型分别有2.8%和4.1%的精度提升；在公开的Allen Human Brain Atlas数据集上相比于SyN算法和VoxelMorph-Diff模型分别有9.5%和5.3%的精度提升。

   The macaque is a crucial experimental animal that shares many genetic, brain organizational, and behavioral characteristics with humans. A macaque brain atlas that identifies anatomically and functionally distinct regions is fundamental to biomedical and evolutionary research. Joint multimodal and cross-scale utilization of medical imaging and bioimaging data can not only provide mesoscopic connectivity and multi-omics mapping, but also facilitate the development of translational medicine, cross-species comparisons and computational modeling, and enrich the collaborative resource platform for non-human primates. The study of spatial registration methods for multi-scale data of macaque brain atlas and indeed the evaluation guidelines for multi-scale integration are important for producing a standard reference space for multi-scale integration of macaque brain atlas. The study of registration methods between bioimaging and medical imaging can establish the correspondence between bioimaging and medical imaging, based on which spatially discrete data with 3D spatial coordinate reference can be embedded into the standard space for subsequent integration of various reference standards to take full advantage of the different modal data for further joint analysis of the brain. A large number of existing registration methods mainly target the unimodal neural system, while few cross-modal and cross-scale registration methods can be applied to brain mapping data with significant differences in grayscale, tissue and anatomical structure, and large differences in image characteristics between data, and different kinds of brain mapping data have both scale and modality differences. There are a series of challenges in integrating the bioimaging and medical imaging of macaque brain atlas through image registration methods.

    In this paper, we systematically study the registration method of integration between rhesus macaque brain atlas bioimaging and medical imaging data, and give the selection criteria of the standard space integration and the measurement index of the integration effect; take the R04 macaque brain imaging data collected by our center as the main data set, use the block face section data in the data set as the intermediate modality and the standard space for integration, and further the Nissl staining data and MRI data of the macaque brain were integrated into the standard space where the block face sections were located. In addition, the proposed method is validated with the publicly available Allen Human Brain Atlas dataset. The main points of this paper are as follows:

    1.Multimodal cross-scale macaque brain image quality control and data processing.This paper first introduces the experimental acquisition process of R04  macaque brain imaging data, and the data are screened through quality control to retain usable data for subsequent experiments; for the unavailable data, the quality problems of the data are investigated through causality analysis to find the causes of the data quality problems in the experimental process, and to assist the experimental instructor to improve the experiments for the purpose of improving the quality of the next dataset. For the unavailable data, we will help the experimental teacher to improve the experiment to improve the quality of the next data set. A complete set of pre-processing methods for Nissl staining data is constructed to address the current situation that there is no unified pre-processing procedure for Nissl staining data; for the usable data after quality control, they are pre-processed into a format that can be directly read into neural networks and medical open source toolboxes.

    2. A new method for multimodal cross-scale registration of macaque brain atlases. The criteria for selecting the standard space are determined, and a set of multimodal cross-scale registration schemes are designed to achieve multimodal registration of macaque brain bioimaging data and medical imaging data using the SyN algorithm for the two main problems of brain atlas data integration: 3D reconstruction and spatial mapping between different scales of data, using R04 macaque brain imaging dataset, We also present a mapping method between different modal data at different scales to solve the initial registration problem of dual different scales. The DSC of regional anatomical label overlap and the Id Score based on L1 loss measure are proposed as the evaluation indexes for the consistency of forward and backward transformation.

     3. Deep learning-based multimodal cross-scale primate brain atlas registration method. In order to improve the processing speed and spatial registration accuracy, the key steps of the multimodal cross-scale registration scheme of macaque brain atlas based on SyN algorithm are improved; a linear registration network and a nonlinear registration framework based on alternate optimization of image migration and registration are proposed respectively; image migration is introduced into cross-modal registration to solve the problem of difficult design of cross-modal registration loss function; circular consistency is introduced into the forward and inverse transform consistency problem The problem of the difference between the image and the original image after sequential forward and inverse transformation is solved to reduce the error of spatial mapping. In the case of guaranteeing the accuracy, we can process a large amount of data quickly. Finally, a multimodal cross-scale macaque brain atlas registration scheme with both speed and accuracy is obtained. The experimental results show that the method proposed under this framework has a 2.8% and 4.1% improvement in accuracy compared with the SyN algorithm and VoxelMorph-Diff model, respectively, on the R04 macaque brain imaging dataset; and has a 9.5% and 5.3% improvement in accuracy compared with the SyN algorithm and VoxelMorph-Diff model, respectively, on the publicly available Allen Human Brain Atlas dataset.