[论文解读] Assessing inter-modal and inter-regional dependencies in prodromal Alzheimer's disease using multimodal MRI/PET and Gaussian graphical models

Martin Dyrba, Reza Mohammadi+3 more|arXiv (Cornell University)|Mar 30, 2018

Functional Brain Connectivity Studies参考文献 67被引用 2

一句话总结

本研究利用高斯图形模型（GGMs）分析667名ADNI受试者在轻度认知障碍、痴呆和健康对照组中，基于多模态MRI/PET数据的淀粉样蛋白β、葡萄糖代谢和灰质体积在不同脑区及不同成像模态间的相互依赖关系。研究发现，不同成像模态间的条件依赖关系极少，网络拓扑结构呈现双相U型轨迹，聚类系数与路径长度变化趋势表明存在模态特异性脑区网络，且GGMs在无需二值化处理的情况下可作为聚类工具使用。

ABSTRACT

Several neuroimaging markers have been established for the early diagnosis of Alzheimer's disease, among them amyloid-beta deposition, glucose metabolism, and gray matter volume. Up to now, these imaging modalities were mostly analyzed separately from each other, and little is known about the regional interrelation and dependency of these markers. Gaussian graphical models (GGMs) are able to estimate the conditional dependency between many individual random variables. We applied GGMs for studying the inter-regional associations and dependencies between multimodal imaging markers in prodromal Alzheimer's disease. Data from N=667 subjects with mild cognitive impairment, dementia, and cognitively healthy controls were obtained from the ADNI. Mean amyloid load, glucose metabolism, and gray matter volume was calculated for each brain region. GGMs were estimated using a Bayesian framework and for each individual diagnosis, graph-theoretical statistics were calculated to determine structural changes associated with disease severity. Highly inter-correlated regions, e.g. adjacent regions in the same lobes, formed distinct clusters but included only regions within the same imaging modality. Hardly any associations were found between different modalities, indicating almost no conditional dependency of brain regions across modalities when considering the covariance explained by all other regions. Network measures clustering coefficient and path length were significantly altered across diagnostic groups, with a biphasic u-shape trajectory. GGMs showed almost no conditional dependencies between modalities when at the same time considering various other regions within the same modalities. However, this approach could be used as a clustering method to derive graph statistics in future studies omitting the need to binarize the network as currently being done for connections based on Pearson correlation.

研究动机与目标

• 探究轻度认知障碍阶段阿尔茨海默病中关键神经影像标志物在不同脑区间及模态间的相互依赖关系。
• 评估这些依赖关系在不同诊断组中随疾病严重程度的变化情况。
• 评估高斯图形模型（GGMs）在无需对连接网络进行二值化处理的情况下，捕捉条件依赖关系的实用性。
• 确定多模态影像标志物（淀粉样蛋白、代谢、萎缩）在不同脑区间是否存在显著的条件关联。

提出的方法

• 从667名ADNI受试者（轻度认知障碍、痴呆、健康对照）的多模态MRI/PET数据中提取淀粉样蛋白负荷、葡萄糖代谢和灰质体积的区域平均值。
• 采用贝叶斯框架估计高斯图形模型（GGMs），推断脑区间在控制所有其他区域后所表现出的条件依赖关系。
• 针对每个诊断组计算图论统计量，包括聚类系数和路径长度，以评估网络拓扑结构的变化。
• 通过检验偏相关系数来评估条件依赖关系，从而隔离其他脑区未解释的关联。
• 该方法通过使用连续的偏相关估计值作为图统计量的基础，避免了对连接网络进行二值化处理。
• 在不同诊断组中分析网络变化，以识别与疾病进展相关的轨迹。

实验结果

研究问题

• RQ1 在轻度认知障碍阶段阿尔茨海默病中，淀粉样蛋白、葡萄糖代谢和灰质体积在不同脑区间的区域相互依赖关系如何变化？
• RQ2 在控制所有其他脑区后，不同神经影像模态之间是否存在显著的条件依赖关系？
• RQ3 图论网络属性（如聚类系数和路径长度）在轻度认知障碍阶段阿尔茨海默病中随疾病严重程度如何变化？
• RQ4 GGMs是否可作为无需连接阈值二值化处理的脑网络聚类方法？
• RQ5 多模态脑影像标志物在不同诊断阶段的网络重组模式是什么？

主要发现

• 高度相关的脑区形成了聚类，但这些聚类主要局限于同一成像模态内部，表明存在模态特异性的区域网络。
• 在不同成像模态之间观察到极少的条件依赖关系，表明当控制所有其他脑区时，某一模态的区域变化并不依赖于其他模态的变化。
• 聚类系数与路径长度在不同诊断组间呈现双相U型轨迹，表明网络变化具有非线性特征，随疾病进展而演变。
• GGM方法成功捕捉了网络拓扑结构，且无需对连接关系进行二值化处理，为传统基于相关性的网络方法提供了更细致的替代方案。
• 区域关联在相同模态内部最强，同一脑叶内相邻区域表现出最高的条件依赖性。
• 结果表明，多模态影像标志物在轻度认知障碍阶段阿尔茨海默病中可能通过主要独立的区域通路发挥作用。