The Development of Individual Differences in Adolescent Brain Structure and Risk

在北卡罗来纳大学早期脑发育中挽救遗漏的纵向MRI访问 研究数据库 项目摘要 在我们正在进行的R01(MH123747-01A1)《青少年大脑中个体差异的发展》中 结构和风险“)项目中，我们的目标是描述大脑结构中个体差异的部分 青春期早期的大脑在生命的早期就已经存在了。青春期早期和青春期是主要的 出生后大脑发育的时期，特征是大脑结构和功能的动态成熟和 重组和新出现的精神障碍风险，尽管尚不清楚这一发展时期是如何 导致大脑结构和风险的个体差异。北卡罗来纳大学早期脑发育研究(EBDS) 是一项独特而创新的纵向研究，跟踪调查了产前登记的儿童，进行了成像和 出生时、1岁、2岁、4岁、6岁、8岁和10岁时的认知/行为评估。这一群体中的482名儿童现在 进入青春期，我们在12岁、14岁和16岁时通过MRI、认知和 行为评估，重点关注执行功能、注意力和焦虑的表型，一致 RDoC的结构对精神障碍风险很重要。一个特别的目标是调查使用 机器学习(ML)用于早期大脑发育对认知和行为的预测分析 这对青春期的结果和随后的精神疾病风险都有影响。然而，大多数机器学习(ML) 当数据点丢失时，应用于纵向数据的算法不能很好地执行(或根本不执行)，如ML 方法需要完整的数据和较大的样本量。因为纵向研究通常会受到 由于收购失败和参与者流失，不同时间点的重大数据丢失，甚至 像UNC EBDS这样的丰富数据库通过仅选择完整可大大降低样本大小 应用预测性ML的数据集(不到1-10岁EBDS数据集的三分之一是完整的)。 在这里，我们建议通过以下方式修复丢失的EBDS时间点(年龄1-10年)的结构性MR图像数据 多模式、多时间点图像预测。该图像数据推算包括跨通道图像 生成(同时从现有MRI数据生成丢失的MRI数据)，以及 纵向数据的多时间点归因(从不同时间点的现有MRI数据生成缺失的MRI数据 时间点)。然后，我们将应用我们的分发外模型来提供有关 推算数据的适当性。随后，应用于原始图像处理的相同图像处理 EBDS MRI数据将应用于输入/生成的MRI数据，以计算 形态测量(区域体积、皮质厚度、表面积和白质纤维束 属性)。这些输入的数据将是对大脑的纵向ML/AI研究非常重要的资源 对EBDS数据集进行了开发，因为这将使培训数据增加200%以上。 原始MR图像、输入的MR图像和形态测量都将通过NDA共享， 与训练有素的归责网络并驾齐驱，供他人使用。

Rescuing Missed Longitudinal MRI visits in the UNC Early Brain Development Studies Database PROJECT ABSTRACT In our ongoing R01 (MH123747-01A1) “The Development of Individual Differences in Adolescent Brain Structure and Risk”) project, we aim to characterize the portion of individual differences in brain structure in the early adolescent brain is already present in the earlier years of life. Early adolescence and puberty is a major period of postnatal brain development, characterized by dynamic structural and functional brain maturation and reorganization, and emerging risk for psychiatric disorders, though it is not known how this period of development contributes to individual differences in brain structure and risk. The UNC Early Brain Development Study (EBDS) is a unique and innovative longitudinal study that has followed children, enrolled prenatally, with imaging and cognitive/behavioral assessments at birth, 1, 2, 4, 6, 8, and 10 years. 482 children from this cohort are now reaching adolescence, and we are following these children at 12, 14, and 16 years of age via MRI, cognitive and behavioral assessments, with a focus on the phenotypes of executive function, attention, and anxiety, consistent with RDoC constructs important for psychiatric disorder risk. One particular aim is to investigate the use of machine learning (ML) for the predictive analysis of early brain development to cognitive and behavioral outcomes in adolescence and to risk for subsequent psychiatric disorders. Yet, most machine learning (ML) algorithms applied to longitudinal data do not perform well (or at all) when data points are missing, as ML methods need both complete data and large sample sizes. As longitudinal studies suffer commonly from significant missing data at different time points due to acquisition failure as well as participant attrition, even a rich database like the UNC EBDS is reduced to a significantly lower sample size by selecting only complete datasets to apply predictive ML (less than a third of the datasets of EBDS data from age 1 – 10 years is complete). Here, we propose to rescue missing EBDS timepoints (at ages 1 - 10 yrs) of structural MR image data via multi-modal, multi-timepoints image predictions. This image data imputation includes cross-modality image generation (generating missing MRI data from existing MRI data at the same time), where available, as well as multi-timepoints imputation of longitudinal data (generating missing MRI data from existing MRI data at different time points). We will then apply our out-of-distribution model to provide additional information on the appropriateness of the imputed data. Subsequently, the same image processing that was applied to the original EBDS MRI data will be applied to the imputed/generated MRI data to compute missing information of morphometric measures (regional volumes, cortical thickness, surface area, and white matter fiber tract properties). This imputed data will be a highly significant resource for longitudinal ML/AI studies of brain development performed on the EBDS dataset, as it would allow for an increase in training data of over 200%. The original MR images, the imputed MR images, and the morphometric measures will all be shared via NDA, alongside the trained imputation network for use by others.