Fast, Robust Analysis of Large Population Data

描述（由申请人提供）：现代成像，如MRI，可以提供全脑的安全、非侵入性测量，并且越来越多地用于脑发育、成熟和衰老的大型临床和研究，以及用于监测药物干预随时间的影响。这就迫切需要开发高度自动化、准确和强大的测量工具来分析大型神经图像数据集。图像配准作为一种重要的图像测量工具已经引起了巨大的科学兴趣，因为它是整合和比较来自不同个体或群体的数据的关键步骤，也是开发反映个体群体内结构和功能变异性的统计图谱的关键步骤。然而，大多数当前的配准算法是基于单个大脑与选定模板的成对配准。这种独立的成对配准和模板的主观选择可以将系统配准误差和偏差引入到对准的图像，从而降低检测细微的大脑变化的统计能力，例如，微小的纵向结构和功能变化，这对早期发现阿尔茨海默病（AD）很重要。为了解决这些局限性，最近提出了组配准和组间比较方法，通过将所有个体受试者直接同步配准到其组平均值，来实现所有受试者的一致配准。然而，这些逐组配准方法的准确性和鲁棒性在识别微小的大脑差异方面受到限制，因为直接从每个受试者到组平均值的潜在大的复杂变形的独立估计可以使最初非常相似的图像（具有微小差异）在配准之后由于配准中的噪声和不确定性而变得非常不同。此外，由于需要同时配准大量图像和计算机存储器容量的限制，当前的分组配准方法只能处理少量图像，例如，几个到几十个。该项目的第一个目标是开发一种快速，鲁棒和准确的分组配准算法，该算法能够同时处理大量图像，例如，成百上千的图像，通过一台普通的计算机。我们的主要思想是将大规模的分组配准问题划分为一系列分层的小规模配准问题，每个问题都可以由通用计算机有效地处理，并且可以通过简化配准问题来鲁棒和准确地解决。 此外，为了有效地比较两个（或多个）组，即，分别从早期疾病患者和正常对照，或从遗传上相同的双胞胎，我们进一步提出了一种新的组间配准方法，有效地对齐两个群体，不仅匹配他们的平均值，而且还匹配他们在所有相应位置的统计分布。因此，两组之间的统计学差异可以很大程度上确定，这使得能够检测由于疾病（例如在AD早期发现的疾病）或双胞胎中微小的大脑生长差异而导致的微小大脑萎缩。这种组间配准和比较方法也可以扩展到多组配准，并应用于新生儿早期双胞胎的纵向研究。所有这些新的组间配准和比较方法的研究是第二个目标的主题。 最后，我们将应用我们开发的组配准方法，以及组间配准和比较方法，ADNI（阿尔茨海默病神经影像学倡议）数据集，用于早期检测AD，以及新生儿数据集，用于研究双胞胎中微小的大脑生长差异。所提出的方法的性能将被广泛验证，并与成对配准方法以及其他组配准方法所获得的性能进行比较。这些研究是第三个目标的主题。 最终开发的算法将通过NITRC（神经成像信息学工具和资源交换所）免费提供给整个研究社区，就像我们的HAMMER注册算法（http：//www.nitrc.org/projects/hammer/）一样，这是NITRC中下载量最大的工具之一。 公共卫生相关性：该项目旨在开发，测试和评估快速，稳健和准确的组配准和统计比较算法，以有效地同时处理大量脑图像;以检测微小，复杂的组差异。这对于脑疾病的早期检测是重要的（例如，阿尔茨海默病）和识别基因相同的双胞胎中微小的大脑发育差异。最终开发的算法将通过NITRC（神经成像信息学工具和资源交换所）免费提供给整个研究社区，就像我们对HAMMER注册算法所做的那样（http：//www.nitrc.org/projects/hammer/），这是目前NITRC中最热门的下载工具之一。

DESCRIPTION (provided by applicant): Modern imaging, such as MRI, can provide a safe, non-invasive measurement of the whole brain, and has been increasingly employed for large clinical and research studies of brain development, maturation, and aging, as well as for monitoring the effects of pharmacological interventions over time. This has created a great need for the development of highly automated, accurate, and robust measurement tools for analysis of large neuroimage dataset. Image registration as an important image measurement tool has attracted enormous scientific interest, since it is the key step for integration and comparison of data from different individuals or groups, as well as for the development of statistical atlases that reflect structural and functional variability within a group of individuals. However, most of the current registration algorithms are based on pair-wise registration of an individual brain with a selected template. This independent pair-wise registration and the subjective selection of template can introduce systematic registration error and bias to the aligned images, thus reducing the statistical power in detecting subtle brain changes, e.g., tiny longitudinal structural and functional changes which are important for early detection of Alzheimer's Disease (AD). To resolve these limitations, group-wise registration and inter-group comparison methods have been recently proposed to achieve consistent registration across all subjects by simultaneous registration of all individual subjects to their group mean directly. However, the accuracy and robustness of these group-wise registration methods are limited in identifying tiny brain differences, since the independent estimation of potentially large complex deformations from each subject to the group mean directly can make the initially very similar images (with tiny difference) become very different after registration, due to noise and uncertainty in the registration. Moreover, because of the required simultaneous registration of a large set of images and the limitation of computer memory capability, current group-wise registration methods can handle only a small number of images, e.g., several to dozens. The first aim of this project is to develop a fast, robust, and accurate group-wise registration algorithm which is able to handle simultaneously a large set of images, e.g., hundreds or thousands of images, by a general computer. Our key idea is to partition a large-scale group-wise registration problem into a series of hierarchical small-scale registration problems, each of which can be handled efficiently by a general computer and can be solved robustly and accurately by simplification of the registration problem. Moreover, for effective comparison of two (or more) groups, i.e., obtained respectively from early-stage diseased patients and normal controls, or from genetically identical twins, we further propose a novel inter- group registration method to effectively align two groups by matching not only their means but also their statistical distributions at all corresponding locations. Thus, the statistical difference between the two groups can be greatly identified, which enables the detection of tiny brain atrophies due to diseases such as those found during the early stage of AD or tiny brain growth differences within twins. This inter-group registration and comparison method can also be extended for the registration of multiple groups, with application in longitudinal study of twins at early neonatal stage. The study of all these novel inter-group registration and comparison methods is the topic of the second aim. Finally, we will apply our developed group-wise registration method, as well as the inter-group registration and comparison method, to the ADNI (Alzheimer's Disease Neuroimaging Initiative) dataset for early detection of AD, and to the neonatal dataset for study of tiny brain growth differences within twins. The performance of the proposed method will be extensively validated and also compared with those obtained by pair-wise registration methods as well as by other group-wise registration methods. These studies are the topic of the third aim. The final developed algorithms will be made freely available to the whole research community through NITRC (Neuroimaging Informatics Tools and Resources Clearinghouse), as we did with our HAMMER registration algorithm (http://www.nitrc.org/projects/hammer/), which is one of the top downloaded tools in NITRC. PUBLIC HEALTH RELEVANCE: This project aims at the development, testing, and evaluation of fast, robust, and accurate group registration and statistical comparison algorithms for effective simultaneous processing of large sets of brain images; to enable the detection of tiny, complex group differences. This is important for early detection of brain diseases (e.g., Alzheimer's Disease) and for identification of tiny brain growth differences within genetically identical twins. The final developed algorithms will be made freely available to the whole research community through NITRC (Neuroimaging Informatics Tools and Resources Clearinghouse), as we did with our HAMMER registration algorithm (http://www.nitrc.org/projects/hammer/), which is currently one of the top download tools in NITRC.