毕业生知识库

脑图谱在神经科学研究及临床医学中扮演着关键角色。然而，由于群体水平脑图谱忽略了个体间脑区差异，在精准医学快速发展的当下，其适用性受到限制。与此相对地，个体水平脑图谱能够提供精确的个体级别脑功能区定位，有效弥补了群体水平图谱的不足，被认为具有广泛的应用前景。近年来，研究者们提出了多种基于磁共振影像（magnetic resonance imaging，MRI）的个体脑图谱绘制方法，但大多数研究仅限于利用单一模态 MRI 所获取的个体信息。相比之下，多模态融合能更全面地捕捉和描述脑功能与结构组织。针对这一现状，本论文对现有的个体脑图谱绘制方法进行了系统梳理，将其分为基于优化的方法和基于学习的方法两大类，并基于这两类中多模态方法存在的局限性展开了针对性研究。本文分别提出了基于优化的和基于学习的多模态个体脑图谱绘制新方法，并从多个维度对所提出方法进行了验证和初步应用研究。本文的主要贡献和创新点包括：

（1）针对现有基于优化的多模态方法在个体图谱同源性方面的不足，本文 提 出 了一种基于多模态脑连接信息的 脑 图 谱 个 体化方 法（multimodal connectivity-based individual parcellation，MCIP）。MCIP 基于弥散磁共振成像（diffusion MRI，dMRI）和静息功能磁共振成像（resting-state functional MRI， rsfMRI），通过融合个体的功能连接和解剖连接特征，优化了脑区内的同质性、空间连续性和与参考图谱的相似性。MCIP 不依赖于多被试数据，可直接生成单个被试的个体图谱，且具有良好的脑区同源性，特别适合应用于疾病患者、非人灵长类动物等小样本数据场景。本研究在个体图谱的功能和解剖同质性、认知行为的预测性能、遗传性、可重复性和跨物种的鲁棒性方面对 MCIP 进行了验证，证实了其在总体上优于现有的基于多模态信息的个体化方法。本研究还对人类和猕猴的脑区拓扑变异进行了比较研究，发现人类的脑区拓扑变异性高于猕猴。总体而言，MCIP 提供了脑区在个体水平上的准确定位，并有潜力用于神经科学和精准医疗中。

（2）任务功能磁共振成像（task functional MRI，tfMRI）反映大脑在执行任务下的功能活动，是一种对个体分区有重要作用的数据模态，然而在科研和临床环境中获取一系列高质量的 tfMRI 数据耗时费力。本文提出了一种基于学习的脑图谱个体化方法（atlas individualizing with task contrasts synthesis，TS-AI），可以利用 tfMRI 对比图的合成帮助个体图谱的绘制。另一方面，通过引入特征一致性损失，TS-AI 克服了现有基于学习的方法因缺乏真实标签而带来的局限性，能充分利用深层神经网络具有的强大学习能力以得到性能更优的个体图谱。本文使用多种参考图谱和数据集对个体图谱的可重复性、同质性和认知行为预测能力进行了评估，展示了 TS-AI 的优越性能和泛化能力。对 TS-AI 进行的敏感性分析揭示了影响各脑区个体差异的多模态特征。此外，TS-AI 能够识别出在阿尔茨海默病进展过程中颞叶内侧和额叶区域的加速萎缩，表明其能够识别疾病生物标记，具有临床研究和应用价值。

综上所述，本研究提出了两种基于多模态磁共振影像的个体脑图谱绘制方法，克服了以往基于多模态信息的个体图谱绘制方法存在的不足，提供了脑区在个体水平上的准确定位，为神经科学和精准医学的发展提供了新工具。

The brain atlas plays a crucial role in neuroscience research and clinical medicine. However, due to the fact that group-level brain atlases ignore individual differences in brain regions, their applicability is limited in the rapidly developing era of precision medicine. In contrast, individual-level brain atlases can provide precise localization of brain functional regions at the individual level, effectively compensating for the inadequacies of group-level atlases and are considered to have broad application prospects. In recent years, researchers have proposed various methods for constructing individual brain atlases based on magnetic resonance imaging (MRI), but most studies have been limited to utilizing individual information obtained from a single modality of MRI. In comparison, fusing multimodal MRIs can more comprehensively capture and describe brain function and structural organization. In response to this situation and fact, this thesis systematically reviewed existing methods for constructing individual brain atlases, dividing them into two main categories: optimization-based methods and learning-based methods, and conducted targeted research based on the limitations of multimodal methods in these two categories. We proposed new optimization-based and learning-based multimodal individual brain atlas construction methods, respectively, and validated and conducted preliminary application studies on the proposed methods from multiple dimensions. The main contributions and innovations of this thesis include:

(1) To solve the problem that the individual atlases obtained by existing optimizationbased multimodal methods lack inter-subject parcel homology, this thesis proposes a Multimodal Connectivity-based Individual Parcellation method (MCIP). MCIP is based on diffusion MRI (dMRI) and resting-state functional MRI (rsfMRI), and optimizes the homogeneity within brain regions, spatial continuity, and similarity with reference atlases by integrating individual functional and anatomical connectivity features. MCIP does not rely on multi-subject data and can directly generate individual atlases with homologous parcels for a single subject, making it particularly suitable for application in small sample data scenarios such as disease patients and non-human primates. We validated MCIP in terms of functional and anatomical homogeneity of individual atlases, predictive performance of cognitive behavior, heritability, repeatability, and crossspecies robustness, demonstrating its overall superiority to existing multimodal individualization methods. We also conducted an exemplary comparative study of topographic variability of parcels in humans and macaques, finding that the topographic variability of humans is higher than that of macaques. Overall, MCIP provides accurate localization of parcels at the individual level and shows potential in neuroscience and precision medicine.

(2) Task functional MRI (tfMRI) reflects brain functional activity during task execution and thus is a modality that plays an important role in individual parcellation. However, acquiring a series of high-quality tfMRI data is time-consuming and laborintensive in research and clinical settings. This thesis proposes a learning-based method, Atlas Individualizing with Task contrasts Synthesis (TS-AI), which can individualize an atlas with the help of synthesizing tfMRI activation maps. On the other hand, by introducing feature consistency loss, TS-AI overcomes the limitations of existing learning-based methods due to the lack of ground truth labels and can fully utilize the powerful learning ability of deep neural networks to obtain better-performing individual atlases. We evaluated the repeatability, homogeneity, and cognitive behavior prediction ability of individual atlases using multiple reference atlases and datasets, demonstrating the superior performance and generalizability of TS-AI. Sensitivity analysis of TS-AI revealed the multimodal features that influence individual differences in each brain region. Furthermore, TS-AI was able to identify accelerated atrophy in the medial temporal and frontal regions during the progression of Alzheimer's disease, indicating its ability to identify disease biomarkers and its value in clinical research and application.

In summary, the two multimodal individual brain parcellation methods proposed in this thesis overcome the shortcomings of previous multimodal individualization methods, provide accurate localization of parcels at the individual level, and offer new tools for the development of neuroscience and precision medicine.