Infant Functional Connectome Fingerprinting based on Deep Learning

项目摘要 功能连接体指纹是发现可靠和健壮的个性化功能连接 能够准确地将一个人与其他人区分开来的图案，如“指纹”。到目前为止， 人们广泛观察到功能性连接体的指纹识别能力，从年龄较大的儿童到 从青少年到成年人。与此同时，对指纹识别贡献最大的功能连接是一致的 被认为是对认知表现最具预测性的。然而，功能连接组指纹图谱 在婴儿期，最具活力的出生后大脑发育仍然没有得到研究，这是 对于理解早期个人层面的职能组织的内在模式至关重要，关系 具有不同行为表型的个体间可区分性，以及相关的异常模式 与产前药物接触有关。婴幼儿功能性连接体研究的两大障碍 指纹：1)在精确处理婴儿神经图像方面存在重大挑战，通常表现为 极低对比度，动态影像表现，形态和功能改变；2)常规 功能连接组指纹图谱的方法简单地使用线性变换的低阶泛函 连通性特征，因此无法分离本质上纠结的身份相关个性化 婴儿大脑中的信息和与年龄相关的发育信息。填补这两种方法中的关键空白 和知识，该项目旨在为婴儿功能开发一种创新的专用深度学习模式 连接体指纹，从而解决了神经发育中的三个基本问题：1) 个性化的功能连接体指纹存在于大脑发育的早期；2)哪些功能 在婴儿期，连接对指纹识别的贡献更大；3)婴儿功能与 有认知能力和不良产前药物暴露的连接体指纹。我们队打得很好 定位于执行此项目，因为我们在开发婴儿专用产品方面拥有丰富的经验 计算工具和深度学习技术，并已获得多个纵向婴儿数据集 既涉及典型的发育中婴儿，也涉及产前接触药物的婴儿。两个具体目标是 建议。在目标1中，我们将开发一个用于婴儿功能连接体指纹识别的深度神经网络模型。 具体地说，为了增强功能连通性特征的区分能力，我们将开发一个三元组 自动编码器模型将这些特征映射到具有高阶判别信息的新的特征空间。 为了抑制来自发展信息的干扰，我们将从潜在变量中分离出来 自动编码器分为身份码、年龄码和噪声码三重，同时设计多个特定损耗 以强制解除纠缠。在目标2中，我们将探索指纹识别和 它们与认知表现和不良产前药物暴露之间的关系。我们的计算模型， 代码和发现将向公众公布，以极大地推进婴儿大脑连接体的研究。

Project Abstract Functional connectome fingerprinting is to discover the reliable and robust individualized functional connectivity patterns that are capable of accurately distinguishing one individual from others, like the “fingerprint”. To date, the fingerprinting capability of functional connectome has been widely observed from older children to adolescents to adults. Meanwhile, the most contributive functional connections for fingerprinting are consistently identified as the most predictive ones for cognitive performance. However, functional connectome fingerprinting during infancy featuring the most dynamic postnatal brain development remains uninvestigated, which is essential for understanding the early individual-level intrinsic patterns of functional organization, the relationship of inter-individual distinguishability with distinct behavioral phenotypes, as well as aberrant patterns associated with prenatal drug exposure. Two major obstacles prevent from investigation of infant functional connectome fingerprint: 1) there exist significant challenges in precisely processing infant neuroimages, which typically exhibit extremely low contrast, dynamic imaging appearance, morphological and functional changes; 2) conventional methods for functional connectome fingerprinting simply use the linearly-transformed, low-order functional connectivity features and are thus unable to separate the intrinsically-entangled identity-related individualized information and age-related developmental information in infant brains. To fill critical gaps in both methodology and knowledge, this project aims to develop an innovative dedicated deep learning model for infant functional connectome fingerprinting, thus addressing three fundamental questions in neurodevelopment: 1) whether the individualized functional connectome fingerprint exists during early brain development; 2) which functional connections contribute more to fingerprinting during infancy; 3) what is the association of infant functional connectome fingerprint with cognitive performance and adverse prenatal drug exposure. Our team is well positioned to conduct this project, as we have extensive experiences in developing infant-dedicated computational tools and deep learning techniques and have acquired multiple longitudinal infant datasets involving both typically developing infants and infants with prenatal drug exposure. Two specific aims are proposed. In Aim 1, we will develop a deep neural network model for infant functional connectome fingerprinting. Specifically, to boost the discriminative capability of the functional connectivity features, we will develop a triplet autoencoder model to map these features into a new feature space with high-order discriminative information. To restrain the interference from the developmental information, we will disentangle the latent variables from the triple autoencoder into identity-code, age-code, and noise-code, and meanwhile design multiple specific losses to enforce the disentanglement. In Aim 2, we will explore the key contributive connections for fingerprinting and their association with cognition performance and adverse prenatal drug exposure. Our computational models, codes and discoveries will be released to public to greatly advance baby brain connectome studies.