Deep Generative Models of Fetal Brain Development: forward modelling of the mechanisms of neurodevelopmental impairment

胎儿大脑发育的深层生成模型：神经发育障碍机制的正向建模

• 批准号: 2741200
• 金额: --
• 依托单位: King's College London
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2022
• 资助国家: 英国
• 起止时间: 2022 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2741200
• 关键词: Deep; Generative; Models; Fetal; Brain

Aim of the PhD Project:Develop novel techniques for surface-to-volume image registration consolidating concepts from discrete and deep optimisation Incorporate biomechanical and biophysical models of brain growth and folding Build generative forward models of neurodevelopment from fetal and neonatal MRI Investigate extensive opportunities for clinical translation following preterm birth, congenital heart disease and epilepsy Project description:Over the period of late gestation, the fetal brain undergoes a period of rapid development. Neuronal cells make distant connections, laying down a network of communication that will later support complex cognition. The rapid growth of cells in the cortex, the outer layer of the brain, creates biomechanical tensions which cause the surface to fold. Disruptions to this process, for example resulting from preterm birth or prenatal conditions such as congenital heart disease, can result in long-term neurodevelopmental impairments such as Autism, ADHE and epilepsy. The objective of this project is to build a forward model of this process, through which pathways of brain injury may be simulated, and the causes of neurodevelopmental impairment be understood. This project extends from previous work that built average models of brain growth using a combination of image registration and Gaussian process regression [1]. This model showed strong potential for the modelling the deviation of preterm development from healthy, at the centre of the brain. But struggled to model the cortex, where considerable variation of individual's brain shape and patterns of cortical organisation make comparison across scans much more difficult. The work also incorporates ideas from [2,3] which built a deep generative model for image-to-image translation and used it to derive feature attribution (FA) maps that highlight all evidence of pathology in individual brains. This is achieved by learning a mapping that changes an image backwards from categorised as diseased towards being classified as healthy. Neither of these projects presented a forward or mechanistic model of disease; however, in [4,5] we propose a novel biomechanical models of brain atrophy following dementia. This supports simulation of the trajectory of brain atrophy under different disease states or conditions: healthy ageing, mild cognitive impairment or full Alzheimer's disease. Accordingly in this project we seek to integrate these ideas to develop a deep generative biomechanical model of cortical growth from late gestation to birth. This will extend the ideas of [4,5] integrating novel models of cortical folding [5] and techniques from the domain of cortical surface registration [6,7] to learn to model longitudinal mappings between an individual's fetal and neonatal scans. Then will adapt ideas from image-to-image models for phenotype translation [2,3] to change tissue contrast and appearance. In this way simultaneously simulating changes in shape, size and tissue maturation.

博士项目的目的：开发新的表面到体积图像配准技术，从离散和深度优化中巩固概念，结合大脑生长和折叠的生物力学和生物物理模型，从胎儿和新生儿MRI中建立神经发育的生成前向模型，研究早产、先天性心脏病和癫痫后临床转化的广泛机会。神经元细胞进行远距离连接，建立一个通信网络，稍后将支持复杂的认知。大脑皮层（大脑的外层）细胞的快速生长产生了生物力学张力，导致表面折叠。这一过程的中断，例如由早产或先天性心脏病等产前条件引起的中断，可能导致长期神经发育障碍，如自闭症，ADHE和癫痫。本项目的目标是建立一个这一过程的正向模型，通过该模型可以模拟脑损伤的途径，并了解神经发育障碍的原因。该项目扩展了以前的工作，即使用图像配准和高斯过程回归的组合建立大脑生长的平均模型[1]。该模型显示出在大脑中心对早产发育与健康的偏差进行建模的强大潜力。但是很难对大脑皮层进行建模，因为个体大脑形状和大脑皮层组织模式的巨大差异使得扫描之间的比较更加困难。这项工作还结合了[2，3]的思想，该思想建立了一个用于图像到图像翻译的深度生成模型，并使用它来导出特征归因（FA）图，突出了个体大脑中病理学的所有证据。这是通过学习一种映射来实现的，该映射将图像从被分类为患病向后改变为被分类为健康。这些项目都没有提出疾病的前瞻性或机械模型;然而，在[4，5]中，我们提出了痴呆后脑萎缩的新生物力学模型。这支持在不同疾病状态或条件下模拟脑萎缩的轨迹：健康老龄化，轻度认知障碍或完全阿尔茨海默病。因此，在这个项目中，我们试图整合这些想法，以开发一个从妊娠后期到出生的皮质生长的深层生成生物力学模型。这将扩展[4，5]的想法，整合皮质折叠的新模型[5]和皮质表面配准领域的技术[6，7]，以学习对个体胎儿和新生儿扫描之间的纵向映射进行建模。然后将从图像到图像模型的想法适用于表型转换[2，3]，以改变组织对比度和外观。以这种方式同时模拟形状、大小和组织成熟的变化。