CASIA OpenIR  > 脑网络组研究
基于磁共振成像的个体化脑图谱绘制及方法研究}
张瀚天
2019-05-31
页数64
学位类型硕士
中文摘要

    脑图谱是研究脑科学重要的工具,它能够帮助我们理解人脑功能和结构的关系,也能为临床诊断提供巨大帮助。从最早的布罗德曼图谱开始,一直到近些年来磁共振技术的进步,越来越多准确、精细的脑图谱出现了。然而,大多数脑图谱的绘制往往注重图谱在人群中的可靠性和准确性,在一定程度上忽视了个体间的差异。另外,由于个体数据有限且个体脑图谱缺乏金标准,个体水平脑图谱的精度仍然有限。针对这一现状,本文系统地研究了基于解剖连接的个体化脑图谱绘制方法,并且给出了对个体化图谱多角度的定量评估方案。本文的主要内容如下:

    本文基于磁共振成像技术,实现了一套基于解剖连接指纹的个体化分区流程。该流程借鉴了传统的解剖连接分区流程,利用基于surface的概率性纤维跟踪算法来获取顶点的解剖连接指纹,以此作为根据来对脑进行分区。我们利用脑网络组图谱提供的数据和信息作为先验知识,将原本没有金标准的个体分区问题转变成了典型的有监督学习问题。我们选取了LightGBM,一种基于梯度提升树的算法框架来作为区分每个亚区的区域性分类器。区域性分类器通过学习大量高质量的被试数据来学习到右脑每个脑区的解剖连接指纹,并用于新来个体的分区。我们发现区域性分类器达到了很高的分区精度,有足够强的能力区分不同的脑区。尽管模型在训练时需要耗费不少时间,但已训练的区域性分类器可以直接投入到个体分区的应用场景,经过平滑的后处理后就能高效准确地完成最终的个体化脑图谱绘制。
    由于个体的分区结果不存在金标准,评估个体化分区质量一直是一个难题。因此,本文希望建立一个全面客观的评估方案,来验证个体化脑图谱的可靠性和准确性。本文使用了39个被试两次采集的磁共振数据对个体化分区流程进行了测试。我们对39个被试的右脑皮层进行分区,发现分区结果与组水平图谱有较好的一致性。更重要的是,个体化的分区结果可以表现出个体的差异性。我们定量计算了被试间差异性在不同脑区的分布,结果符合现在的研究,初级皮层的脑区相对于联络皮层在人群中更为稳定。最后,由于不同的解剖连接可以表征不同的脑区功能,我们计算了脑区的解剖连接指纹相似性,发现绘制的个体化脑图谱比组水平的图谱更能反映每个脑区的功能特异性。

英文摘要

    Human atlas is an important tool for understanding brain. It can help us understand the relationship between  brain function and structure.It is also helpful for clinical. With the advancement of magnetic resonance technology in recent years, more and more fine-grained brain atlas have been established. However, most atlas tends to focus on the reliability  of the parcellations in the population while  ignores the impact of individual differences on brain regions.On the other hand, due to the limited individual data and the lack of gold standard for individual parcellation, the precision of individual-level atlas is limited. Aiming at this situation, this thesis systematically investigated the method of individual-level brain mapping based on structural connectivity.It also gives a quantitative evaluation scheme for various aspects of subject-specific atlas. The main contents of this paper are as follows:

   We performed  individual parcellation method based on structural connectivity with diffusion tensor imaging. The pipeline drawed on the traditional connectivity-based parcellation method and employed the surface-based probabilistic tracking algorithm to obtain the structural connectivity fingerprint of the vertex, which was used as a basis to parcellate the brain. With no standard-gold for individual brain atlas, we combined the data and information provided by the Brainnetome atlas to transform unsupervised problem into a typical supervised learning problem. We chose LightGBM, an algorithmic framework based on gradient boost trees, as areal classifiers that distinguished each subregion.We trained areal classifiers with high-quality data to learn the structural connectivity fingerprint of each brain region, and employed trained areal classifier for parcellations of new comes. We found that regional classifiers achieve high accuracy and have enough discriminate ability between different brain regions. Although our model consumed a lot of time to train, the trained regional classifier can directly invest in the application scenario of individual parcellation with smooth post-processing.The subject-specific brain atlas was obtained efficiently and accurately with areal classifiers.

    As individual-level atlas has no  gold standard, assessing the quality of individualized parcellations has always been a problem. Therefore, this paper hopes to establish a comprehensive evaluation pipeline to validate the reliability and robustness of the individual-level atlas. In this paper, the results of the magnetic resonance data collected by 39 subjects with two sessions were tested. We found that our parcellation results were reproducible while retained  differences cross subjects. We also calculated the spatial distribution of inter-subject variability in different brain regions. We found that the results were consistent with intuition that the primary cortex is more stable than association cortex. In addition, dissimilarity of structural connectivity of  the brain regions was computed. We found that the individual-level brain maps  were more representative of the functional specificity of each brain region than the group-level maps.

关键词磁共振成像,个体化脑图谱,有监督学习,纤维跟踪,扩散张量成像
语种中文
七大方向——子方向分类脑网络分析
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/23853
专题脑网络组研究
推荐引用方式
GB/T 7714
张瀚天. 基于磁共振成像的个体化脑图谱绘制及方法研究}[D]. 中国科学院自动化所. 中国科学院自动化所,2019.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
张瀚天_个体化分区硕士论文.pdf(13840KB)学位论文 开放获取CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[张瀚天]的文章
百度学术
百度学术中相似的文章
[张瀚天]的文章
必应学术
必应学术中相似的文章
[张瀚天]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。