Statistical Methods for Mapping Human Brain Development

描述（由申请人提供）：现代神经成像技术将发育神经科学带入了一个前所未有的突破时代的门槛。通过对越来越多的儿童和成人样本进行高分辨率脑部扫描，研究人员正在绘制出大脑在整个生命周期中的正常发育，以及与精神疾病相关的发育异常。这些研究通常需要在数以万计的大脑位置拟合模型，部分由于这种高计算负荷，研究人员倾向于采用次优方法。一个突出的例子是拟合皮质厚度随年龄发展的多项式模型。在拟合单个模型时，非参数平滑提供了众所周知的优于多项式依赖的优势，但到目前为止，平滑方法尚未应用于同时拟合数千个模型的设置。更广泛地说，迫切需要最先进的统计方法来处理由发育中的大脑研究产生的大量神经成像数据集。本提案的目的是为正常和异常大脑发育的统计分析提供一个全面的工具包。研究人员已经开始为此开发一些创新技术，并建立了一个强大的多机构合作关系，理想地适合于迎接未来的许多挑战。第一个具体目标集中在计算上可行的大量曲线的估计，这些曲线表示一个感兴趣的数量的分布的平均值或给定的百分位数；

DESCRIPTION (provided by applicant): Modern neuroimaging technology has brought developmental neuroscience to the threshold of an era of unprecedented breakthroughs. With high-resolution brain scans acquired in increasingly large samples of children and adults, investigators are mapping both the normal development of the brain over the lifespan, and the developmental abnormalities that are associated with psychiatric disorders. These studies typically entail fitting models at tens of thousands of brain locations, and due in part to this hih computational load, investigators have tended to settle for suboptimal methods. A prominent example is fitting a polynomial model for the development of cortical thickness with age. Nonparametric smoothing offers well-known advantages over polynomial dependence when fitting a single model, but to date, smoothing methodology has not found application to settings in which many thousands of models are fitted concurrently. More broadly, there is a critical need for state-of-the-art statistical methods to tackle the massive neuroimaging data sets generated by studies of the developing brain. The objective of this proposal is to provide a comprehensive toolkit for statistical analyses of normal and abnormal brain development. The investigators have begun to develop a number of innovative techniques toward this end, and have forged a strong multi-institution collaboration ideally suited to meeting the many challenges that lie ahead. The first specific aim focuses on computationally feasible estimation of large numbers of curves representing the mean, or a given percentile, of the distribution of a quantity of interest, conditional on a predictor such as age. The second aim encompasses several hypothesis testing methods that are particularly relevant to neuroimaging, including tests of polynomial null hypotheses against smooth alternatives, as well as tests for group differences in developmental trajectories and other complex outcomes. The third aim, originally motivated by the need for succinct visual representations of spline fits at each point in a grid of brain locations, is to develop novel methods for clustering large amounts of functional data. The proposed methods will be applied to data acquired by multiple imaging modalities, including resting-state functional magnetic resonance imaging, diffusion tensor imaging, and cortical thickness measurement. Most of the methods proposed here are applicable to any imaging modality, and many can be applied outside the field of neuroimaging. Thus the proposed research will have a significant impact both on statistical methodology and on neuroscience, psychiatry, and other disciplines.