Personalized Functional Network Modeling to Characterize and Predict Psychopathology in Youth

摘要 固有功能连接性磁共振成像是研究脑组织结构的有力工具 人脑的功能网络(FN)。大量和不断积累的证据表明，FNS 在青年中进行可预测的规范性发展，而异常发育与 不同的精神病态。基于图像分析方面的进展的最新工作已经确定，FN在 因人而异的事实。当与大规模神经成像数据集配对时，特定于个人的FN提供 前所未有的转换机会，为开发新的诊断技术提供了可能 神经精神疾病的个性化治疗。然而，特定于人的翻译承诺 FNS目前受到几个障碍的阻碍。首先，当前方法在特定的位置计算个性化的FN 规模，尽管有明确的证据表明，大脑是一个具有层级功能组织的多尺度系统。 其次，为了加强不同主题之间的对应关系，个性化的FN通常在以下项下计算 某些限制，这可能会产生有偏见的结果。第三，深度学习在以下领域取得了喜忧参半的成就 神经成像数据分析部分是因为通常采用自组织网络架构，并且 通过采用预先设计的功能而不是学习的功能，功能学习能力经常被剥夺。第四，到 来自大规模神经成像研究的多个数据集的神经成像测量的正确部位效应 目前的方法通常试图在统计建模之前协调数据，导致有价值的损失 信息。第五，纵向神经影像和临床数据越来越多，但有效的分析 用于纵向数据的工具很少。最后但并非最不重要的一点是，深度学习算法已经发展到 分析fcMRI数据，但通常以文档记录不佳的源代码形式发布，这限制了可重复性和 被翻译研究人员采用。在这项申请中，我们在前一个获奖期的成功基础上， 通过开发、验证和传播表征大脑功能的工具来解决这些限制 个人学科层面的组织。我们将利用对大脑的补充大规模研究 开发以验证我们的方法，并描述FNS的异常发育如何与 青少年精神病理学的主要方面，包括抑郁、焦虑、精神病和ADHD谱系 症状。具体地说，我们将开发新的方法来1)准确地识别无偏见的个性化FN 多尺度的层级组织；2)使用个性化的方法稳健地预测精神症状维度 具有优化的深度神经网络结构和集成的现场效应校正的FNS，以及3)有效 对FNS的纵向数据进行建模，以创建精神病理学的预测模型。这些工具将于年发布 免费提供的集装箱化软件包，可确保计算平台之间的无缝移植 和完全的重复性。

ABSTRACT Intrinsic functional connectivity magnetic resonance imaging is a powerful tool to study the organization of functional networks (FNs) in the human brain. Rich and accumulating evidence demonstrates that FNs undergo predictable normative development in youth, and that abnormal development is associated with diverse psychopathology. Recent work based on advances in image analytics has established that FNs are in fact person-specific. When paired with large-scale neuroimaging datasets, person-specific FNs provide unprecedented translational opportunities for the development of new diagnostics that could guide personalized treatments for neuropsychiatric illnesses. However, the translational promise of person-specific FNs is at present hindered by several obstacles. First, current methods compute personalized FNs at a specific scale, despite clear evidence that the brain is a multi-scale system with a hierarchical functional organization. Second, to enforce correspondence across different subjects personalized FNs are typically computed under certain constraints, which may yield biased results. Third, deep learning has achieved mixed success in neuroimaging data analysis partially due to the fact that ad-hoc network architecture is typically adopted and feature learning capability is often deprived by adopting pre-engineered rather than learned features. Fourth, to correct site effects of neuroimaging measures from multiple datasets of large-scale neuroimaging studies current methods typically attempt to harmonize data prior to statistical modeling, resulting in loss of valuable information. Fifth, longitudinal neuroimaging and clinical data are increasingly available, but effective analytic tools for longitudinal data are scarce. Last but not least, deep learning algorithms have been developed to analyze fcMRI data but are often released as poorly documented source code, limiting both reproducibility and adoption by translational researchers. In this application, we build on the success of the prior award period to address these limitations by developing, validating, and disseminating tools that characterize brain functional organization at an individual subject level. We will leverage complementary large-scale studies of brain development to validate our methods and delineate how abnormal development of FNs is associated with major dimensions of psychopathology in youth, including depression, anxiety, psychosis, and ADHD-spectrum symptoms. Specifically, we will develop novel methods to 1) accurately identify bias-free personalized FNs with a multiscale hierarchical organization; 2) robustly predict psychiatric symptom dimensions using personalized FNs with optimized deep neural network architecture and integrated site-effect correction, and 3) effectively model longitudinal data of FNs to create predictive models of psychopathology. These tools will be released in a freely available, containerized software package to ensure frictionless portability across computing platforms and full reproducibility.