Quantifying Brain Abnormality by Multimodality Neuroimage Analysis

通过多模态神经图像分析量化大脑异常

• 批准号: 8964568
• 负责人: Dinggang Shen
• 金额: $ 36.57万
• 依托单位: UNIV OF NORTH CAROLINA CHAPEL HILL
• 依托单位国家: 美国
• 财政年份: 2012
• 资助国家: 美国
• 起止时间: 2012-08-01 至 2020-03-31
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/8964568
• 关键词: Address; Adopted; Algorithms; Alzheimer' s Disease; Amyloid; Appearance; Architecture; Atlases; Brain; Categories; Clinical; Communities; Complex; Computer software; Computers; Country; Coupled; Data; Diagnosis; Diagnostic; Discipline of Nuclear Medicine; Disease; Early Diagnosis; Elderly; Functional disorder; Future; Goals; Heterogeneity; Image; Individual; Joints; Learning; Machine Learning; Maps; Methods; Mining; Modality; Modeling; Molecular; Nature; Nerve Degeneration; Neurodegenerative Disorders; Neurologist; Neuropsychological Tests; Pathology; Pattern; Preventive Intervention; Research; Sampling; Scientist; Selection Bias; Software Tools; Staging; Structure; Symptoms; Techniques; Technology; Therapeutic; Time; Training; Work; comparison group; design; diagnostic accuracy; disease diagnosis; forest; imaging modality; improved; in vivo; multimodality; multitask; neuroimaging; novel; outcome forecast; pre-clinical; public health relevance; success; vector

 DESCRIPTION (provided by applicant): Alzheimer's disease (AD) develops for an unknown and variable amount of time before its symptoms fully manifest. But, when the symptoms become clinically observable, a significant neurodegeneration has already taken place. Thus, there is a largely unmet need for technologies that can aid the effective early diagnosis and prognosis of AD in an in vivo and more objective manner. The goal of this renewal project is to develop a set of advanced machine-learning techniques for precise in vivo quantification of pathological changes of brains with multimodality neuroimaging for both early diagnosis and prognosis of AD. AD is a highly heterogeneous neurodegenerative disorder with complex pathophysiology, thus very challenging to pinpoint its subtle pathologies without any aid from advanced computational technologies. To this end, we propose the following four specific aims to identify those subtle disease-induced alterations, derive robust diagnostic conclusions, and predict future disease trajectories. Specifically, in Aim 1, we will develop a multi-view feature representation technique to robustly extract complementary information from neuroimaging data with multiple representative atlases, and then identify a small subset of most discriminative features for AD diagnosis. This novel multi-atlas technique will deviate from the conventional single-atlas approaches in feature representation, which are often susceptible to inter-subject structural variability, registration error, and atlas selection bias. In Aim 2, we will further devlop two novel multi-view feature mapping techniques for collaborative fusion of multimodality information by explicitly considering the distribution heterogeneity of different categories of features extracted from different modalities. This will significantly avoid the unnecessary complexity of feature distributions after our collaborative fusion, thus increasing the efficacy of subsequent diagnostic classifiers. Specifically, a deep learning technique (with deep multi-layered architecture) will be adopted to hierarchically mine multimodality information that resides nonlinearly both within each modality and between different modalities. In Aim 3, we will develop a novel multi-task sparse learning technique for joint prediction of diagnostic status and clinical scores (e.g., ADAS-Cog and MMSE) by considering the inherent correlations between features and between training samples. This will also allow us to exploit the latent structure underlying the data for robust estimation of these highly variable clinical scores. Finally, in Aim 4, we will jointly predict clinical scores of each given subject in multiple future time points, by developing coupled random forests that can take advantage of all training subjects with complete or even incomplete multimodality data and further enforce temporal consistency of those estimated clinical scores. All the above-proposed techniques will be evaluated by a large image set of elderly subjects in ADNI. We expect that the successful completion of this renewal project will result in a comprehensive and effective diagnosis/prognosis framework for improving early detection of AD. The respective software tools will be released freely to the research community, as we have done with our HAMMER software, which has been downloaded by >5200 users from >20 countries.

 描述(由申请人提供)：阿尔茨海默病(AD)在其症状完全显现之前，会发展一段未知和可变的时间。但是，当临床上可以观察到这些症状时，已经发生了显著的神经变性。因此，对能够以活体和更客观的方式帮助AD有效的早期诊断和预后的技术的需求在很大程度上尚未得到满足。这一更新项目的目标是开发一套先进的机器学习技术，利用多模式神经成像技术对脑内病变进行精确的在体量化，以用于AD的早期诊断和预后。AD是一种高度异质性的神经退行性疾病，具有复杂的病理生理机制，因此在没有任何先进计算技术的帮助下精确定位其微妙的病理非常具有挑战性。为此，我们提出了以下四个具体目标，以确定这些细微的疾病诱导的变化，得出可靠的诊断结论，并预测未来的疾病轨迹。具体地说，在目标1中，我们将开发一种多视角特征表示技术，从具有多个代表性图谱的神经成像数据中稳健地提取互补信息，然后识别出用于AD诊断的最具区分性的特征的一小部分。这种新颖的多图谱技术在特征表示上将与传统的单一图谱方法相背离，传统的单一图谱方法往往容易受到被试之间的结构差异、配准误差和图谱选择偏差的影响。在目标2中，我们将进一步发展两种新的多视角特征映射技术，通过明确考虑从不同模式提取的不同类别特征的分布异质性，来实现多模式信息的协同融合。这将显著避免协作融合后特征分发的不必要复杂性，从而提高 后续诊断分类器。具体地说，将采用深度学习技术(具有深度多层体系结构)来分层地挖掘驻留的多模式信息 在每种模式内和不同模式之间都是非线性的。在目标3中，我们将开发一种新的多任务稀疏学习技术，用于诊断状态和临床的联合预测 通过考虑特征之间和训练样本之间的内在相关性来获得分数(例如，ADAS-Cog和MMSE)。这也将使我们能够利用潜在的数据结构，对这些高度可变的临床评分进行稳健的估计。最后，在目标4中，我们将 联合预测每个给定受试者在未来多个时间点的临床分数，通过开发 耦合随机森林，可以利用所有训练对象的完整甚至不完整的多模式数据，并进一步加强这些估计的临床分数的时间一致性。以上提出的所有技术都将通过ADNI上的一组老年受试者的图像进行评估。我们期望这一更新项目的成功完成将产生一个全面和有效的诊断/预后框架，以改进AD的早期发现。各个软件工具将免费发布给研究社区，就像我们对HAMMER软件所做的那样，该软件已被来自20个国家和地区的5200名用户下载。