Statistical Learning Methods for Neuroimaging Data Analysis with Applications Hongtu Zhu1 Tengfei Li2 and Bingxin Zhao3

2025-05-03 0 0 2.02MB 75 页 10玖币

侵权投诉

Statistical Learning Methods for Neuroimaging Data

Analysis with Applications

Hongtu Zhu1, Tengfei Li2, and Bingxin Zhao3

1Departments of Biostatistics, Statistics, Genetics, and Computer Science and

Biomedical Research Imaging Center, University of North Carolina, Chapel Hill

2Departments of Radiology and Biomedical Research Imaging Center, University

of North Carolina, Chapel Hill

3Department of Statistics and Data Science, University of Pennsylvania

arXiv:2210.09217v1 [stat.AP] 17 Oct 2022

Abstract

The aim of this paper is to provide a comprehensive review of statistical challenges in

neuroimaging data analysis from neuroimaging techniques to large-scale neuroimaging studies

to statistical learning methods. We brieﬂy review eight popular neuroimaging techniques and

their potential applications in neuroscience research and clinical translation. We delineate

the four common themes of neuroimaging data and review major image processing analysis

methods for processing neuroimaging data at the individual level. We brieﬂy review four

large-scale neuroimaging-related studies and a consortium on imaging genomics and discuss

four common themes of neuroimaging data analysis at the population level. We review nine

major population-based statistical analysis methods and their associated statistical challenges

and present recent progress in statistical methodology to address these challenges.

Keywords: causal pathway, heterogeneity, image processing analysis, neuroimaging

techniques, population-based statistical analysis, study design.

1 Introduction

Neuroimaging refers to the process of producing images of the structure, function, or pharmacol-

ogy of the central nervous system (CNS). It has been a dynamic and evolving ﬁeld with (A1) the

development of new acquisition techniques, (A2) the collection of various neuroimaging data in

clinical settings and medical research, and (A3) the development of statistical learning (SL) meth-

ods. For (A1), popular neuroimaging techniques include structural magnetic resonance imaging

(sMRI), functional magnetic resonance imaging (fMRI), diffusion weighted imaging (DWI), com-

puterized tomography (CT), positron emission tomography (PET), electroencephalography (EEG),

magnetoencephalography (MEG), and functional near-infrared spectroscopy (fNIRS). These tech-

niques were developed to measure speciﬁc tracers in CNS, that are directly and indirectly associ-

ated with brain structure and function. For instance, PET delineates how an injected radioactive

tracer (e.g., Fluorodeoxyglucose (FDG)) moves and accumulates in the brain, whereas fMRI mea-

sures an indirect tracer, called the concentration of deoxyhaemoglobin, in the ﬂow downstream of

the activated neurons caused by brain’s activity. The developments of SL methods for individual

neuroimaging data raise serious challenges for existing statistical methods due to four common

themes consisting of (CT1) complex brain objects,(CT2) complex spatio-temporal structures,

(CT3) extremely high dimensionality, and (CT4) heterogeneity within subjects and across groups.

For (A2), in recent years, huge amounts of neuroimaging data have been collected in health

care, biomedical research studies, and clinical trials. First, neuroimaging has the potential to im-

prove clinical care for diagnosis and prognosis in various brain-related diseases, such as demen-

tia, sleep disorders, and schizophrenia. Some typical uses of neuroimaging include identifying

the effects of brain-related diseases (e.g., stroke or glioblastoma), locating cysts and tumors, and

ﬁnding swelling and bleeding, among others. Second, many large-scale biomedical studies have

collected/are collecting massive neuroimaging data (e.g., sMRI, DWI, and fMRI) with high spa-

tial and/or temporal resolution as well as other complex information (e.g., genomics and health

factors) in order to map the human brain connectome for understanding the pathophysiology of

brain-related disorders, the progress of neuropsychiatric and neurodegenerative disorders, the nor-

mal brain development, and the diagnosis of brain cancer, among others. In the last two decades,

there are at least three pioneering neuroimaging-related studies, including Alzheimer’s Disease

Neuroimaging Initiative (ADNI) (http://www.adni-info.org/) (Weiner et al., 2010), the Human

Connectome Project (HCP) (http://humanconnectome.org/consortia/) (Van Essen et al., 2013), and

the UK Biobank (UKB) study (https://www.ukbiobank.ac.uk/) (Miller et al., 2016). They repre-

sent major advances and innovations in acquisition protocols, analysis pipelines, data management,

experimental design, and sample size. The left panel of Figure 1 shows multi-view data across dif-

ferent domains (e.g., imaging, genetics, or environmental factors) in some large-scale biomedical

studies. Third, neuroimaging biomarkers have many uses in clinical trials for drug development

in neurological and psychiatric disorders (Schwarz, 2021). These uses include a screening tool for

selecting trial participants, a tool to establish biodistribution, target engagement and pharmacody-

namic activity, a means for monitoring safety, and an evidence measure of disease modiﬁcation.

The developments of SL methods for clinical translation and large-scale neuroimaging-related

studies raise serious challenges to existing statistical methods due to the four additional themes of

(CT5) sampling bias,(CT6) complex missing patterns,(CT7) complex data objects, and (CT8)

complicated causal pathways in brain disorders.

For (A3), there is a large literature on the development of SL methods for neuroimaging data

analysis (NDA) in order to correlate multi-type data from different domains across multiple stud-

ies, eventually establishing a dynamic causal pathway (e.g., the causal genetic-imaging-clinical

(CGIC) pathway in the right panel of Figure 1) linking genetics to brain (or neuroimaging) pheno-

types to clinical outcomes confounded with health factors. These SL methods can be categorized

into two categories including image processing analysis (IPA) at individual level and population-

based statistical analysis (PSA) for a sample of subjects. We further group various IPA methods

into deconvolution and structure learning (Sotiras et al., 2013; Li et al., 2019; Zhou et al., 2021;

Shen et al., 2017; Park et al., 2003; Yi et al., 2019). Deconvolution methods primarily include

image reconstruction and image enhancement. Structure learning methods mainly include image

segmentation and image registration. Due to (CT1)-(CT4) and the lack of high-quality annotation

datasets, it is very challenging to develop ‘good’ IPA pipelines to extract a relatively small number

of image phenotypes (IPs) with high repeatability and reproducibility for both individual health

care and PSA. We also group various PSA methods into nine main categories, including study de-

sign, statistical parametric mapping, object oriented data analysis, dimensional reduction methods,

data integration, imputation methods, predictive models, imaging genetics, and causal discovery

(Ombao et al., 2016; Nathoo et al., 2019; Shen and Thompson, 2019; Smith and Nichols, 2018;

Nichols et al., 2017; Rathore et al., 2017). Due to (CT1)-(CT8), each category has its own statistical

challenges, requiring speciﬁc statistical methodology to address them. However, the development

of scalable PSA methods has fallen seriously behind the technological advances in neuroimaging

techniques, causing difﬁculty in translating research ﬁndings to clinical practice.

Environment

Genomics

Imaging

Cognition Disease

Unseen

Confounder

Unseen

Confounder

Unseen

Confounder

Behavior

Genetics

Environment

Lifestyle

SES

Imaging

Confounders

Observed

Unobserved

Clinical

Disease Cognition Behavior

UKB 500000

ABCD 10000+

HCP Lifespan 4000+

ADNI 2000+ …

Figure 1: Left panel: Major data types from different domains in several representative large-scale

biomedical studies; Right panel: A dynamic causal model for delineating causal genetic-imaging-

clinical (CGIC) pathway confounded with environmental factors and unobserved confounders.

2 A Review of Neuroimaging Techniques and Uses

We brieﬂy review eight neuroimaging techniques below and summarize them in supplementary

Table 1. For each image modality, we describe its tracer, data dimension, features, main uses,

and several key softwares (Smith and Webb, 2010). Figure 2 also presents different neuroimaging

modalities and the different types of features they extract.

•Structural magnetic resonance imaging (sMRI) measures the ﬂuid characteristics of different

tissues (gray and white matter), creating high-resolution (0.5mm-1mm) images with a strong

gray/white matter contrast and many anatomical details. It allows us to qualitatively and

quantitatively measure the development and change of cortical and subcortical structures in

terms of both size and shape in the brain. Some sMRI derived measurements include cortical

thickness, cortical folding, sulcal depth, voxel-based morphometry, and regional volumes

and shape. sMRI has been widely used for diagnosis, staging, and follow-up of disease in

clinics and brain development in research.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

StatisticalLearningMethodsforNeuroimagingDataAnalysiswithApplicationsHongtuZhu1,TengfeiLi2,andBingxinZhao31DepartmentsofBiostatistics,Statistics,Genetics,andComputerScienceandBiomedicalResearchImagingCenter,UniversityofNorthCarolina,ChapelHill2DepartmentsofRadiologyandBiomedicalResearchImagingCenter...

展开>> 收起<<

Statistical Learning Methods for Neuroimaging Data Analysis with Applications Hongtu Zhu1 Tengfei Li2 and Bingxin Zhao3.pdf

共75页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Statistical Learning Methods for Neuroimaging Data Analysis with Applications Hongtu Zhu1 Tengfei Li2 and Bingxin Zhao3

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: