Individualised modelling of dynamic brain networks in fMRI.

大脑是一个复杂的器官。功能性磁共振成像（fMRI）是一种强大的非侵入性工具来记录人类的大脑活动。它检测血氧水平依赖性（BOLD）信号作为神经活动的间接测量，在不同大脑区域具有高空间分辨率。为了研究这些区域在大脑网络中如何相互作用，科学家们计算了功能连接，在fMRI中，功能连接是来自不同区域的BOLD信号之间的相关性。神经科学家已经确定了在休息和任务中相互协调的大脑网络。传统上，使用静态方法来估计功能连接性-使用整个长扫描会话来计算每对大脑区域的单个相关值。最近，人们对理解大脑网络动力学越来越感兴趣，即，功能连接的动态性。大脑活动预计会一直变化，因此时变网络描述可以捕获静态方法错过的信息。这一领域的另一个重要趋势是考虑受试者的变异性。例如，功能性大脑区域的空间位置可以随受试者而变化。通常，通过将具有相似活动的体素分组在一起来识别功能性大脑区域。然后，为大脑区域提取单个时间序列，并将其作为数据馈送到动态方法，例如隐马尔可夫模型（HMM）。获得这些大脑区域的时间过程之前已经完成了使用组级独立成分分析（伊卡）结合二元回归，以给出每个主题的功能性大脑区域的主题特定的版本。然而，这已被证明是表现不佳相比，新开发的方法，如PROFUMO，明确模型的主体变异性的生成模型。考虑到受试者的变异性对于理解健康和患病的人类大脑至关重要;因为，它提供了针对不同类型大脑的更具体的描述，而不仅仅是一个总体模糊的群体描述。这对于在大数据时代提供个性化医疗至关重要。在过去的十年中，我们见证了大规模公开可用的神经影像数据的发展，如人类连接组计划（HCP）和英国生物银行（UKB），并提出了一个巨大的挑战：我们如何通过使用来自大队列的信息来理解个体受试者大脑网络的功能和故障？本博士项目从动态脑网络的角度迎接挑战。我们的目标是在功能磁共振成像中模拟具有特定主题时空变异性的网络动态。首先，我们将开发强大且值得信赖的指标来验证不同的动态模型，例如隐马尔可夫模型（HMM）和动态网络模式（Dynemo）。我们将探索贝叶斯推理度量（如变分自由能）和机器学习方法（如交叉验证），以衡量不同模型可靠地表示大脑网络动态的能力。这将为我们未来的工作提供模型和超参数选择的信息。其次，我们将探索网络动力学的遗传基础。以前的工作已经发现，动力学是高度遗传的HCP数据中使用的孪生结构。我们将通过探索从英国生物库数据中现有的动态网络模型获得的单核苷酸多态性和表型之间的关联来超越这项工作。最后，我们将建立一个新的动态网络模型，更好地处理主题的可变性。与以往的方法不同，该模型将是端到端的-将这两个步骤结合成一个大模型，并对100，000个UKB数据的主题进行特定描述。该项目福尔斯EPSRC医学成像研究领域-包括医学图像和视觉计算。工业主管是罗氏公司的Stanislaw Adaszeewski博士。

The brain is a complex organ. Functional magnetic resonance imaging (fMRI) is a powerful non-invasive tool to record brain activity in humans. It detects the blood-oxygen-level-dependent (BOLD) signal as an indirect measure of neural activity, with high spatial resolution across different brain regions. To study how these regions interact with each other in brain networks, scientists calculate functional connectivity, which in fMRI is the correlation between BOLD signals from distinct regions. Neuroscientists have identified brain networks that coordinate with each other in both rest and task. Traditionally, functional connectivity is estimated using a static approach - calculating a single correlation value for each pair of brain regions using the entirety of a long scanning session. Recently, there has been increasing interest in understanding brain network dynamics, i.e., dynamics in functional connectivity. Brain activity is expected to be changing all the time, and hence time-varying network descriptions can capture information being missed by static approaches. Another important trend in this field is to consider subject variability. For example, the spatial locations of functional brain regions can vary over subjects. Typically, functional brain regions are identified by grouping together voxels with similar activity. A single timeseries is then extracted for the brain region and fed in as data to the dynamic approaches such as the Hidden Markov Modelling (HMM). Obtaining these brain region time courses has previously been done using group-level Independent Component Analysis (ICA) combined with dual regression to give a subject-specific version of each subject's functional brain region. However, this has been shown to underperform compared to newly developed methods such as PROFUMO, which explicitly model subject variability in a generative model. Accounting for subject variability is critically important for understanding the human brain in both health and diseased; as, rather than just one overall vague group description, it provides more specific descriptions tuned to different types of brains. This is crucial for the delivery of personalised medicine in the era of big data. The last decade has witnessed the development of large-scale publicly available neuroimaging data, such as the Human Connectome Project (HCP) and UK Biobank (UKB), and posed a great challenge: how do we understand functioning and malfunctioning of individual subject brain networks by using information from the large cohort? This DPhil project faces up to the challenge from the perspective of dynamic brain networks. We aim to model network dynamics with subject-specific spatiotemporal variability in fMRI. First, we will develop robust and trustworthy metrics to validate the different dynamic models such as the Hidden Markov Modelling (HMM) and Dynamic Network Modes (Dynemo). We will explore Bayesian inference metrics (such as variational free energy) and machine learning methods (such as cross validation) to measure the ability of different models to reliably represent the brain network dynamics. This will inform model and hyperparameter selection for our future work. Second, we are going to explore the genetic basis of network dynamics. Previous work has found that dynamics are highly heritable using twin structure in HCP data. We will go beyond this work by exploring associations between single nucleotide polymorphisms and phenotypes obtained from existing dynamic network models in UK Biobank data. Finally, we will build a new dynamic network model that better handles subject variability. Different from previous methods, this model will be end-to-end - combining these two steps into one large model with subject-specific descriptions on 100,000 subjects of UKB data.This project falls within the EPSRC research area of medical imaging - including medical image and vision computing. The industrial supervisor is Dr. Stanislaw Adaszeewski from Roche.