Deep Generative Analyses for fMRI data

fMRI 数据的深度生成分析

• 批准号: 10820636
• 负责人: Daniela Frank De Albuquerque
• 金额: $ 4.05万
• 依托单位: DUKE UNIVERSITY
• 依托单位国家: 美国
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-11-01 至 2025-10-31
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10820636
• 关键词: Address; Adult; American; Behavior; Biological; Biological Markers; Brain; Brain region; Code; Cognitive; Computer software; Cost of Illness; Data; Data Analyses; Data Set; Dependence; Disease; Emotions; Experimental Designs; Explosion; Functional Magnetic Resonance Imaging; Functional disorder; Genetic Markers; Goals; Health; High Prevalence; Impairment; Individual; Individual Differences; Knowledge; Learning; Link; Major Depressive Disorder; Maps; Mental disorders; Methods; Modality; Modeling; Morphology; Neurosciences; Neurosciences Research; Outcome; Participant; Patients; Pattern; Persons; Prevention; Process; Psychiatric therapeutic procedure; Psychiatrist; Psychiatry; Research Personnel; Series; Severities; Structure; Sum; Symptoms; System; Time; Training; Translating; Work; associated symptom; autoencoder; biomarker identification; brain abnormalities; brain based; brain morphology; brain volume; care delivery; cognitive function; cohort; comparative; cost; design; emotion regulation; experimental study; flexibility; high dimensionality; improved; individual variation; interest; lost earning; machine learning model; multimodal neuroimaging; neural network; neuroimaging; neuropsychiatric disorder; novel; personalized health care; personalized intervention; recurrent neural network; response; severe mental illness; social; sound; spatiotemporal; stem; tool; treatment response; vector

ABSTRACT About 51.5 million people (1 in 5 US adults) lives with a mental illness (MI) and it is estimated that serious MI costs Americans about $193 billion in lost earnings, yearly. Given the high prevalence and social cost of MI, there has been a growing push for translating advances in neuroscience research into improvements in MI prevention and psychiatry care delivery. In this context, it has become increasingly evident that psychiatric diseases emerge as result of abnormalities in brain spatiotemporal dynamics and network connectivity. Furthermore, neuropsychiatric diseases typically have a high degree of individual variability in presentation, symptom severity, and treatment response. In this proposal, we aim to design new fMRI analysis methods capable of tackling the abovementioned challenges – i.e., capable of directly modeling brain spatiotemporal dynamics, while also capturing individual variability. More specifically, the main goal of this proposal is to extend a previously developed deep-generative fMRI analysis model (VAE-GAM) that produces interpretable spatial effect maps for each covariate (as in standard methods) while capturing nonlinear effects and correlations across voxels. To accomplish this goal, I propose to: 1) Model temporal dynamics directly by fitting a Recurrent Neural Network (RNN) to the VAE-GAM latent space; and 2) Capture individual differences by using a deep Mixed Effects Modeling framework to model individual subject maps as being the sum of a group-level baseline map and a subject-unique map, generated using a learned, subject-unique embedding vector. The expected outcome of this proposal is a flexible fMRI analysis toolset that will allow researchers and clinicians to identify new brain activity patterns linking high-level behavior in health and disease states. We believe such a model could be a step towards fulfilling the goal of delivering biologically-sound, computationally driven, and personalized health care for millions of patients afflicted by mental illness.

摘要 约有5150万人(五分之一的美国成年人)患有精神疾病(MI)，据估计，严重的MI 美国人每年损失的收入约为1930亿美元。鉴于心肌梗塞的高患病率和社会成本， 越来越多的人推动将神经科学研究的进步转化为MI的改善 提供预防和精神科护理服务。在这方面，越来越明显的是，精神病患者 疾病的出现是大脑时空动力学和网络连接异常的结果。 此外，神经精神疾病通常在表现上具有高度的个体差异性， 症状严重程度和治疗反应。在这个提案中，我们的目标是设计新的fmri分析方法。 能够应对上述挑战--即能够直接对大脑进行时空建模 动态，同时也捕捉个体的变异性。更具体地说，这项提案的主要目标是将 以前开发的深度生成功能磁共振分析模型(VAE-GAM)，可以产生可解释的空间 每个协变量的效果图(如在标准方法中)，同时捕获 体素。为了实现这一目标，我建议：1)通过拟合递归神经网络来直接建模时间动力学 网络(RNN)到VAE-GAM潜在空间；以及2)通过使用深度混合来捕捉个体差异 效果建模框架，将单个主题地图建模为组级别基线地图的总和 以及使用学习的、主题唯一的嵌入向量生成的主题唯一地图。预期的结果 这项提议的一个灵活的功能磁共振分析工具包，将允许研究人员和临床医生识别新的大脑 将健康和疾病状态中的高级行为联系起来的活动模式。我们相信这样的模式可能是一种 朝着实现提供生物健康、计算驱动和个性化健康的目标迈进一步 照顾数以百万计的精神疾病患者。