AD classification algorithms using ADNI multi-modal image data

描述（由申请人提供）：最近阿尔茨海默病（AD）研究的一个主要重点是确定AD临床前状态早期干预的特征和途径。 这是至关重要的，因为当患者表现出症状性痴呆时，大脑中的数百万神经元已经丢失。 能够在早期阶段可视化AD相关脑变化的先进成像方法可能在早期检测神经退行性变的发作中发挥重要作用，在广泛的神经损失之前。 同时，必须开发自动化图像分析算法，不仅能够准确定位和表征变化，而且还能够从数据中识别模式，以实现个体受试者诊断。 最近的结果表明，使用机器学习技术确实可以实现这一点：几个小组使用T1加权MR（磁共振），FDG PET（18氟脱氧葡萄糖正电子发射断层扫描）或其他方式获得了约85%的灵敏度和特异性。 ADNI数据中纵向随访受试者的多种成像方式的可用性现在为AD诊断工具的灵敏度/特异性提供了显著改善的机会。 这可以通过开发新的数学模型来实现，以利用多种成像模式和纵向数据结合起来，以进行准确的诊断，以及帮助分析新的治疗方法的效果，因为它们变得可用。 假设：通过使用多种成像模式和纵向数据，可以显著提高区分AD、MCI（轻度认知障碍）和健康对照（个体受试者水平）的准确性（灵敏度/特异性）。 具体目标：（1）开发新的基于图像的分类算法，可以在统一的框架内同时利用多种成像模式;并在整个ADNI图像数据集上广泛评估这些方法。 (2)为了扩展我们的算法，以纳入纵向图像数据，以进一步改进，并对ADNI图像数据进行评估。 (3)设计及推广一个开放源码软件系统，以协助诊断阿尔茨海默病。 研究方法：我们最近开发了一种新的基于机器学习的AD分类算法，该算法使用ADNI的T1加权MR图像从健康对照中正确识别AD患者的灵敏度/特异性为85%，使用FDG PET图像的分类灵敏度/特异性超过80%。 在这些努力的基础上，我们将扩展和改进我们的模型，以利用多模态图像数据（具体目标1）。 该算法将进一步增强纵向数据，以便更准确地预测难以诊断的病例（具体目标2）。 我们将广泛评估我们的方法在ADNI数据集上的逐步改进，并对ADNI中可用的各种认知/实验室生物标志物的分类置信度等指标之间的关系进行统计评估。 这些算法将通过易于使用的软件系统实施和传播，以促进AD诊断（具体目标3）。 公共卫生相关性：拟议的研究将开发新的算法和软件系统，用于高度准确的阿尔茨海默病诊断。 该技术将利用来自两种或更多种成像模式的个体的大脑图像-例如结构图像（例如，磁共振）和功能图像（例如，正电子发射断层扫描）。 该项目成功完成后的算法和软件将联合收割机与患者的纵向数据（如果可用）相结合，为个体受试者的AD诊断提供宝贵的帮助。

DESCRIPTION (provided by applicant): A major emphasis in recent Alzheimer's disease (AD) research is identifying features and avenues for early intervention in the preclinical state of AD. This is essential because by the time a patient exhibits symptomatic dementia, millions of neurons in the brain have already been lost. Advanced imaging methods that enable visualization of AD associated brain changes at an early stage are likely to play an important role in detecting the onset of neurodegeneration early, prior to extensive neural loss. In parallel, automated image analysis algorithms must be developed to enable not only accurate localization and characterization of changes but also identify patterns from the data to enable individual subject diagnosis. Recent results suggest that this may indeed be possible using machine learning techniques: sensitivity and specificity of around 85% have been obtained by several groups using either T1-weighted MR (Magnetic Resonance), FDG PET (18fluorodeoxyglucose Positron Emission Tomography), or other modalities. The availability of multiple imaging modalities for longitudinally followed subjects within the ADNI data now offers the opportunity of significant improvements in sensitivity/specificity of diagnostic tools for AD. This can be achieved by developing new mathematical models to leverage multiple imaging modalities and longitudinal data in conjunction-for accurate diagnosis as well as to aid in analyzing the effect of new treatments, as they become available. Hypothesis: Significant improvements in accuracy (sensitivity/specificity) for discriminating AD, MCI (Mild Cognitive Impairment), and healthy controls (at the level of individual subjects) are possible by making use of multiple imaging modalities and longitudinal data. Specific Aims: (1) To develop new image-based classification algorithms that can take advantage of multiple imaging modalities simultaneously within a unified framework; and to extensively evaluate these methods on the entire ADNI image dataset. (2) To extend our algorithms to incorporate longitudinal image data for further improvements and perform evaluation on ADNI image data. (3) To design and disseminate an open source software system to assist in AD diagnosis. Methods: We have recently developed a new machine learning based algorithm for AD classification that gives 85% sensitivity/specificity in correctly identifying AD patients from healthy controls using T1-weighted MR images from ADNI, and over 80% classification sensitivity/specificity using FDG PET images. Building on these efforts, we will extend and improve our models to utilize multi-modal image data (Specific Aim 1). The algorithms will be further augmented with longitudinal data to enable more accurate prediction for difficult to diagnose cases (Specific Aim 2). We will extensively evaluate the stepwise improvements of our method on the ADNI dataset and perform statistical evaluations of the relationship of measures such as the classification confidence with various cognitive/laboratory biomarkers available within ADNI. These algorithms will be implemented and disseminated via an easy to use software system to facilitate AD diagnosis (Specific Aim 3). PUBLIC HEALTH RELEVANCE: The proposed research will develop novel algorithms and software systems for highly accurate Alzheimer's disease diagnosis. The techniques will take advantage of the individual's brain images from two or more imaging modalities-such as structural images (e.g., Magnetic Resonance) and functional images (e.g., Positron Emission Tomography). The algorithms and software as a result of successful completion of this project will combine the multi-modal image information with the longitudinal data of the patient (if available) to provide invaluable assistance in AD diagnosis at the level of individual subjects.