CASIA OpenIR  > 毕业生  > 博士学位论文
基于磁共振成像的人类脑网络组图谱的绘制及其方法研究
李海
学位类型工学博士
导师蒋田仔
2017-12
学位授予单位中国科学院大学
学位授予地点北京
关键词分区 脑图谱 人类脑网络组图谱 磁共振成像 扩散张量成像 纤维跟踪 并行计算
摘要
    脑图谱是脑科学中十分重要的基础研究工具,准确、精细、可靠的脑图谱将极大促进我们人类的脑科学研究。自一百多年前布罗德曼图谱发布以来,随着相关技术的进步,脑图谱一直处在不断转变和升级当中。最近十多年,利用磁共振成像技术,基于连接信息的脑区分区获得了领域内同行越来越多的重视,而应用该方法进行完整脑图谱绘制的系统研究还非常少。针对这一现状,本论文系统地研究了基于解剖连接信息的脑图谱绘制及其方法。本论文主要内容和贡献如下:
    基于宏观层次的磁共振成像技术,整合并优化了基于解剖连接的脑区分区流程。该流程整合了前人的研究成果,利用概率性纤维跟踪算法来刻画体素/顶点之间的解剖连接,通过分别在个体水平上进行谱聚类分析以及组水平上进行结果对齐整合实现脑区的分区,本文对流程中的多个环节进行了优化,包括组分析中标签匹配算法的改进,谱聚类算法效率的提升等。本文还提出了基于三种准则——分区模式间一致性准则,分区模式内一致性准则和拓扑一致性准则——的多项验证指标,通过这些指标可以确定一个合适的分区类别数。此外,该流程可实现基于volume和基于surface两种方式的脑区分区。为了验证该脑区分区流程的可重复性,我们利用一套单被试高频采样数据,对该流程分区进行测试,发现该流程具有较高的可重复性。
    面对磁共振影像数据量大、计算耗时、操作繁琐等挑战,配套开发了一套能完整实现以上脑区分区流程的开源软件——ATPP。该软件具有自动化、并行化和使用灵活等特点。用户只需提供输入文件,设置好参数,软件便会自动地运行一整套脑区分区流程,并最终生成所需的结果和日志,全程再无需人工干预。该软件能同时提供机器内(多核CPU加速或GPU加速)和跨机器(在高性能计算集群中分配任务)两种级别的并行化支持,能同时处理大批量的磁共振影像数据,可大大加快脑区分区进程。软件既提供了命令行模式,可支持高性能计算集群上的多脑区并行分区,也提供了直观的图形界面模式,可用于个人台式机上的单脑区分区,方便用户调试参数。该软件提供了模块化的功能设计,方便用户根据实际情况定制自己的流程。该软件同时实现了基于volume和基于surface两种方式的分区流程。我们利用ATPP在两套数据集上分别进行了测试验证。此外,ATPP也已在多篇文献中得到了成功的应用和完整的验证。
    利用上述流程和软件,基于一套公开的高质量数据集,借助高性能计算服务器集群,系统地绘制了人类脑网络组图谱,并基于另一套数据集得到了较好的验证。人类脑网络组图谱共包含有246个精细亚区,其中皮层上210个亚区,皮层下结构36个亚区。每个亚区都有各自的概率图、解剖连接模式、功能连接模式以及相应的功能表征。我们制作了一个二值化的解剖连接网络,方便用户了解各亚区之间的连接。同时,我们也制作了surface版本的脑网络组图谱,方便用户进行皮层相关的分析。此外,为了方便用户查看和使用,我们在多个离线和在线的平台上公开发布了人类脑网络组图谱的所有数据。
其他摘要
    Brain atlases are fundamental research tools in brain science, accurate, fine-grained and reliable brain atlases will greatly facilitate brain science. Since the release of Brodmann atlas over a hundred years ago, brain atlases have been evolving and escalating with advances in related technologies. In the last ten years, connectivity-based parcellation based on magnetic resonance imaging has gained more and more attention in the community. However, There is still short of systematic research on brain atlas using these methods. The thesis, in view of this situation, systematically investigated the mapping and its methods of brain atlas based on anatomical connectivity. The main contributions of the thesis are as follows:
    Based on the magnetic resonance imaging technologies at macroscale, we integrated and optimized the workflow of anatomical connectivity-based parcellation. Based on previous research, probabilistic fiber tracking algorithm was used to characterize the anatomical connectivity between voxels/vertices, we parcellated the brain region by combining the spectral clustering analysis at the individual level as well as alignment and integration of the results at group level. Some steps in the workflow are optimized, such as the method of label matching in group level analysis and the eiffiency of clustering algorithm. We proposed multiple indices following three criteria, (1) consistency across parcellations criterion, (2) consistency within parcellations criterion, and (3) consistency of topology criterion, through which the optimal solutiong k can be determined. Besides, the workflow implements volume-base and surface-based parcellation. To validate the reproducibility of the workflow, we tested the workflow based on a single-subject highly sampled dataset and found the high reproducibility of the workflow.
    In the face of large amount of magnetic resonance imaging data, the time-consuming processing and the tedious operation, we developed an open source software, ATPP, that completely realizes the above pipeline for parcellation. The software features automation, parallelization and flexibility. Fed with input files and configurations by users, the software will automatically run the whole parcellation pipeline and ultimately generate the results and logs without the human intervention. The software provides parallel computing both within machine (multi-core CPU acceleration or GPU acceleration) and across machine (dispatch tasks in high-performance computing clusters). It can handle large volumes of magnetic resonance imaging data, thus can greatly accelerate the progress of parcellation. The software provides not only the command-line mode for multi-ROI oriented parcellation in high-performance computing clusters, but also the intuitive graphical user interface mode for single-ROI oriented parcellation and debug on individual desktops. The software provides modular functional design which is convenient for users to customize their own pipeline. Besides, the software implements the volume-base and surface-base parcellation. We tested ATPP based on two independent dataset. Besides, ATPP has been successfully utilized and fully validated in a number of literature.
    Based on the above workflow, software, and a set of open high-qulity data, we systematically mapped the Human Brainnetome Atlas with the help of high-performance computing clusters. The atlas was well validated based on another dataset. The Human Brainnetome Atlas contains a total of 246 subregions, of which 210 locate in cortex and 36 locate in subcortical structure. Each subregion has its own probabilistic map, anatomical connection, functional connection and functional characterization. We created a binarized anatomical connection network that allows the users to better understand the connections between subregions. At the same time, we also produced a surface version of the Human Brainnetome Atlas, which is convenient for users in cortex related analysis. Besides, to promote the visualization and utilization of the Human Brainnetome Atlas, we publicly release our results in a number of offline and online platforms. 
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/15511
专题毕业生_博士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
李海. 基于磁共振成像的人类脑网络组图谱的绘制及其方法研究[D]. 北京. 中国科学院大学,2017.
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
Thesis.pdf(33528KB)学位论文 暂不开放CC BY-NC-SA请求全文
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
[李海]的文章
百度学术
百度学术中相似的文章
[李海]的文章
必应学术
必应学术中相似的文章
[李海]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。