Spatial-extent inference and prediction in brain imaging data

拟议的研究计划侧重于开发统计方法，解决人类大脑磁共振成像（MRI）技术中的一个关键问题。脑成像数据信息丰富，但高维且复杂，表现出空间自相关性，需要统计建模和推理来解开和提高功率。特别是，由于脑成像研究中使用的受试者数量相对较少，因此有必要开发一个强大的统计模型来捕获大多数脑成像数据的变化。尽管最近的MRI技术获得了高分辨率图像，但目前的统计方法并没有充分利用丰富的信息。一个关键的挑战是将空间高斯过程应用于高维数据和排列以控制误报的巨大计算成本。此外，即使计算成本放松，如何将其用于聚类推理也不清楚，这通常用于提高灵敏度。因此，目前的统计实践包括降低采样和平滑（模糊）图像，这是不充分的和功率不足。该计划将开发一种新颖而统一的方法，以解决空间范围推断的统计和计算挑战。拟议的研究计划有三个具体目标。首先，PI将考虑稀疏性通知高斯过程，将其应用于计算多变量检验统计，然后开发一种扩展扫描统计的聚类推理方法和计算效率高的置换方法。其次，PI将通过建模空间自相关来估计多式联运对应和遗传性，将其扩展到脑成像研究。最后，所提出的方法将与高维中介分析相联系，这对于理解与基因型和表型相关的大脑功能和解剖学至关重要。PI将通过大规模脑成像数据库的应用，包括青少年大脑认知发展（ABCD）研究、人类连接组计划（HCP）和英国生物银行，来验证所提出的方法。该计划提出了一种及时的统计方法，将吸引统计学家和从业人员的重大关注。建议的方法将通过可重复性的R包广泛实施。它将为学生学员提供宝贵的机会，并鼓励他们参与从实施到出版的整个研究过程。

The proposed research program focuses on developing statistical methods that address a crucial issue in magnetic resonance imaging (MRI) technology for the human brain. Brain imaging data are rich with information but are high-dimensional and complex, exhibiting spatial autocorrelations that require statistical modelling and inference to disentangle and improve power. In particular, because the number of subjects used in brain imaging studies is relatively small, it is necessary to develop a powerful statistical model to capture most variations of the brain imaging data. Despite high-resolution images obtained by recent MRI technology, current statistical methods do not fully take advantage of the rich information. A key challenge is the massive computational cost of applying the spatial Gaussian process to high dimensional data and permutation to control false positives. Furthermore, even when the computational cost is relaxed, it is unclear how it can be used for clusterwise inference, which is commonly used to improve sensitivity. As a result, current statistical practices include downsampling and smoothing (blurring) the images, which are insufficient and underpowered. This program will develop a novel and unified methodology that addresses statistical and computational challenges for spatial-extent inference. The proposed research program has three specific objectives. First, the PI will consider a sparsity-informed Gaussian process, apply it to compute multivariate test statistics, and then develop a clusterwise inference method that extends scan statistics and a computationally efficient permutation method. Second, the PI will extend it to research in brain imaging by modelling spatial autocorrelation to estimate intermodal correspondence and heritability. Lastly, the proposed approach will be linked to the high-dimensional mediation analysis, which is essential for understanding the brain functions and anatomies related to genotypes and phenotypes. The PI will validate the proposed methodology through applications to large-scale brain imaging databases, including the Adolescent Brain Cognitive Development (ABCD) study, Human Connectome Project (HCP), and UK BioBank. The proposed program addresses a timely statistical methodology that will attract significant attention from both statisticians and practitioners. The proposed methodology will be widely implementable through an R package for reproducibility. It will provide invaluable opportunities for student trainees, and they will be encouraged to participate in the whole process of research, from implementation to publication.