新生儿时期是人类后天发育的起始点。人脑的结构与功能网络在此阶段已经初步形成，为后期的认知与行为能力发展奠定基础。近年来，无创的脑功能磁共振成像技术快速发展，特别是与计算机信息技术的深度融合，使研究者可以基于新生儿群体的脑影像数据，无创地对人脑在出生阶段的宏观功能连接模式进行解析，进而探索人脑功能环路的形成及配置原则。新生儿脑功能研究已经成为理解人脑功能形成乃至智能涌现的重要前沿领域。
对人脑的各类同质神经组织进行归类分区，探索人脑不同尺度的最小功能单元，称为人脑分区图谱。随着计算机科学技术的迅速发展，传统的人脑图谱学已经进入了“数字化”阶段，尤其是功能分区标准图谱的出现，为揭示人脑的基本功能单元提供了全新途径，成为了各类脑影像研究的通用基础工具与先验“知识库”。当前国际主流的人脑功能图谱，大多是基于健康成人群体绘制的。儿童脑并不是成人脑的“缩小版本”，具有其特有的功能特征与描述方法。因此，儿童脑智发育研究需要不同年龄段特有的脑功能分区图谱。然而，由于新生儿脑影像数据在采集与分析上的巨大挑战，现今在国际上新生儿脑功能分区图谱非常匮乏。如何基于无创的脑功能影像数据，提取高质量的功能连接分类特征，绘制新生儿标准脑功能分区图谱，是领域内亟待解决的关键技术问题。
本论文采用人脑发育连接组计划（www.developingconnectome.org）中公开的52名足月新生儿多模态脑影像数据（37-42孕周）进行脑分区图谱绘制。首先，我们计算每个新生儿全脑灰质体素的功能连接矩阵和欧式距离矩阵，得到灰质体素之间的功能距离矩阵，进而进行层级聚类对每个体素进行分类标号。然后，我们采用两个体素在某一个个体分区图谱中被分到同一类的概率，来定义52个个体分区图谱之间体素水平的群体相似性。最后，结合归一化的脑区匀质性与稳定性指标，绘制了群体水平上的新生儿标准脑功能分区图谱。该脑分区图谱包括120个功能分区，相比于常用的随机增长算法与随机分半算法获得分区，均具有显著更高的区域内功能匀质性。为了验证所绘制的脑分区图谱的有效性，论文利用另一组独立的新生儿影像数据集对所绘制的分区图谱与四种现有的新生儿脑分区图谱（三种脑结构分区图谱和一种基于解剖分区边界的脑功能分区图谱），进行了两个方面的评价分析：发现所构建的新生儿标准脑功能分区图谱比已有的脑分区图谱具有最高的功能匀质性；当采用所绘制的新生儿脑分区图谱定义脑网络节点进而提取个体的功能连接网络特征时，发现所绘制的新生儿脑分区图谱比已有脑图谱在两种机器学习预测模型中，均具有最高的婴儿个体脑龄预测准确度。
综上所述，本论文基于新生儿静息态脑功能影像数据，提出了一种基于体素功能连接模式的两步式聚类算法，成功绘制了新生儿标准脑功能分区图谱，在独立样本上的验证分析表明所绘制的新生儿脑功能分区图谱比已有图谱具有更好的分区同质性和脑龄预测能力。该分区图谱能够为研究者在婴幼儿脑发育研究中提供匀质的脑功能基本单元定义，为研究新生儿的脑发育提供了重要的基础性工具，有望被用于基于脑影像的婴幼儿个体发育、个体识别以及脑发育性疾病等相关研究，对于探索人脑功能的早期发育规则具有重要价值。

The neonatal period is the beginning of postnatal neurodevelopment. At this phase, both structural and functional brain networks have already been established, which lay a foundation for cognitive and behavioral abilities in later life. Recent studies have utilized non-invasive functional magnetic resonance imaging (fMRI) in combination with computer science and technology, which allows researchers to acquire high-quality brain images of neonates and further explore the emergence and development of brain functional networks. The study of neonatal brain networks has become an important frontier for understanding early brain development and the emergence of intelligence. 
Classifying the brain into different structural and functional units has been refereed as brain parcellation atlases. With the rapid advances of neuroimaging, the brain parcellation has stepped into a “digital” stage. Specifically, the functional parcellation provides a new insight into functionally organized units in the human brain that become a basic "prior knowledge database" for neuroimaging research. Currently, the existing functional brain parcellations are mostly based on healthy adults. As the neonate brain is not a “small version” of an adult brain, there is a crucial need for neonate-specific functional parcellations to study early brain development. However, due to the huge challenges in the acquisition and analysis of neonatal brain imaging data, the functional parcellation of neonatal brain is currently rare. Therefore, an important and challenging issue is to study how to extract functional connectivity features based on the neonate fMRI data and further develop the function parcellation atlases of neonate brain. 
In this thesis, we propose a group-wise neonatal functional parcellation framework based on a two-level clustering algorithm using resting-state fMRI data of 52 participants (postmenstrual ages from 37 to 42 weeks) from the developing Human Connectome Project. First, for each neonate we calculate individual voxel-wise functional connectivity matrix and Euclidean distance matrix of the whole brain, followed by a hierarchical clustering analysis. This process generates 52 individualized functional parcellations, which are considered as inputs into an N-Cut clustering algorithm at different scales. We then estimate the resolution of our parcellation, and eventually generate a group-wise functional brain parcellation with 120 regions of interests. Furthermore, the homogeneity of our parcellation is compared with that of the null models generated by a random region growth method and a random half-splitting method. Results show that our approach can significantly capture more homogeneous functional subunits than random parcellations. Next, we evaluate functional homogeneity in another independent dataset. We compared the homogeneity of our parcellation with four existing parcellations in brain development studies (3 structural parcellations and 1 functional parcellation). We also construct neonate brain networks whose nodes are defined by each of these parcellations and put the functional connectivity as a feature into two machine learning algorithms to predict the scan age for each neonate. We find that in the independent dataset, our brain parcellation shows significantly higher functional homogeneity and better performance in age prediction tasks, which verifies the validity of our approach.
In summary, we generate a neonate functional parcellation based on brain connectivity features and a two-level clustering algorithm. This proposed parcellation overcomes the existing neonate brain parcellations in both functional homogeneity and age predictions, thus providing a basic tool to study brain development in early life. We anticipate that this parcellation atlas will facilitate the research of neonate brain development, individual differences, individual identification, and developmental disorders.