Integrating large-scale neural mass modeling and deep learning

背景近年来神经成像领域的进展使得利用功能磁共振成像(FMRI)创建基于弥散磁共振成像的结构连接图及其功能对应图成为可能。识别对了解大脑功能至关重要的网络结构及其动力学的重要特征和机制，需要制定方法来揭示大脑结构、功能和神经成像信号背后的特征生物物理机制之间的联系。多尺度建模就是这样一种策略。然而，多尺度建模无法有效地组合来自不同来源和不同分辨率水平的大数据集(例如，EEG/MEG和fMRI数据集)，并且受限于当前神经科学中不断变化的理论理解。另一方面，深度人工神经网络(DEN)被用于对高维神经成像数据进行分类，以建立特定数据特征与临床变量之间的联系。然而，丹斯的方法在嘈杂和不完整的数据上运行，不能提供对神经生理学的机械性见解。这表明，在分析大型神经成像数据集时，多尺度生物物理建模和深度学习可以有效地互补：在深度学习揭示相关性的情况下，生物物理建模可以在较低尺度上将原因分解为机制。首要目标是提出并验证一个大规模的大脑活动神经团模型，该模型集成了一个现实的深度学习算法。目标1.发展一种非对称连接的生物真实性学习方法。目标2.将目标1中发展的学习方法整合到大脑活动的大规模神经团模型中。目标3.使用模拟数据和易于理解的基准数据集对模型进行校准和测试。目的4.从健康受试者的fMRI和EEG/MEG数据中估计参数并预测系统动力学。将深度学习和大规模神经团模型相结合，将使我们能够将大脑结构、功能和神经生理机制的研究结果联系起来。我们的工作可能会改变该领域的传统思维，该领域的传统思维主要集中在细胞层面上整合生物学和机器学习，DANS单位模仿神经元。这里，我们提出了一种平均场方法，该方法侧重于“神经元质量”的计算能力，而不是对网络中的每个神经元进行单独建模。

Background Progress in the neuroimaging field in recent years has made it possible to create diffusion MRI-based structural connectivity maps as well as their functional counterparts, with functional MRI (fMRI). The identification of important features and mechanisms of network structures and their dynamics that are critical to understand brain function have necessitated the formulation of methods to unravel the links between brain structure, function, and characteristic biophysical mechanisms underlying neuroimaging signals. Multiscale modeling is one such strategy. However, multiscale modelling can fail to efficiently combine large datasets from different sources and different levels of resolution (e.g, EEG/MEG and fMRI datasets) and is constrained to current but ever-changing theoretical understandings in neuroscience. On the other hand, deep artificial neural networks (dANNs) are used to classify high-dimensional neuroimaging data for establishing links between specific data features and clinical variables. However, dANNs methods operating on noisy and incomplete data cannot provide mechanistic insights into neurophysiology. This suggests that multiscale biophysical modeling and deep learning can effectively complement each other when analyzing large neuroimaging dataset: where deep learning reveals correlation, biophysical modeling can unpack cause into mechanisms at lower scales. Overarching goal To propose and validate a large-scale neural mass model of brain activity that has integrated a realistic deep learning algorithm. Objective 1. Develop a biologically realistic learning method with asymmetric connections. Objective 2. Integrate the learning method developed in Objective 1 into a large-scale neural mass model of brain activity. Objective 3. Calibrate and test the model with simulated data and well-understood benchmark datasets. Objective 4. Estimate parameters and predict system dynamics from fMRI and EEG/MEG data recorded from healthy human subjects. Integrating deep learning and large-scale neural mass models will allow us to link findings of brain structure, function, and neurophysiological mechanisms. Our work can potentially shift conventional thinking in the field, which has focused mainly on integrating biology and machine learning at the cellular level, with dANNs units modeled after neurons. Here we propose a mean field approach that focuses on the computing capabilities of `neural masses' rather than modeling each neuron in the network individually.