【作者】 唐振宇；

【导师】 苏敬勇；

【作者基本信息】 哈尔滨工业大学 ， 计算机技术（专业学位）， 2025， 硕士

【摘要】 功能磁共振成像(functional Magnetic Resonance Imaging,fMRI)是研究大脑神经活动及其功能网络动态变化的重要工具。fMRI通过测量血氧水平依赖性信号,间接反映神经元活动引起的血流动力学变化,从而推断神经活动的特征。传统的一般线性模型(General Linear Model,GLM)分析方法依赖线性假设,难以全面捕捉复杂神经活动的非线性特性,在脑功能解析中存在显著局限性。流形学习作为一种无监督降维方法,能够在高维数据中提取低维非线性结构,在表征和分析复杂神经活动方面展现出独特的优势。然而,现有流形学习方法在处理个体水平fMRI数据时,未能充分利用fMRI信号的高维空间结构和时间依赖特性,导致适配性不足和特征提取能力受限。为解决这些问题,本文提出了一个新的框架,包括基于流形学习的特征模式提取方法和基于特征模式的区域同步性检测方法,其中前者提供深层表征支持,后者基于提取的特征模式实现任务相关神经活动的高精度检测和解析。本文首先提出了一种基于拉普拉斯特征映射的人脑图特征模式(BGEm-LE)提取方法,用于检测大脑神经活动的深层流形表征。通过对个体化fMRI数据进行预处理与数据增强,并结合流形学习方法的低维嵌入分析,本文探讨了局部线性嵌入和等度量映射在捕捉大脑功能网络结构中的潜力与局限性。分析结果表明,尽管经典方法能够揭示任务态下功能网络的动态变化,但在特征模式分离和任务相关性提取中仍存在不足。在此基础上,本文提出了BGEm-LE方法,通过构建基于组水平二值脑图的个体化带权功能连接图,结合拉普拉斯特征分解提取出一组具有生物学意义的大脑特征模式。实验结果表明,BGEm-LE方法在功能区域划分和特征模式分离上的表现优于局部线性嵌入和局部线性嵌入,并展现出更高的分辨率和生物学相关性。基于BGEm-LE提取的特征模式,本文设计了一种基于图特征模式的区域同步性检测方法(RS-GEm),用于精准定位任务相关激活区域,以检测大脑神经活动。该方法结合提取特征模式的功能领域信息、解剖结构约束和区域同步性测量框架,计算特定脑区内功能邻域的协同活动水平。具体而言,RS-GEm方法通过测地线距离限制功能邻域的计算范围,抑制了远距离伪连接的干扰,从而确保分析结果的生理合理性。此外,该方法通过同步性测量捕捉体素及其功能邻域的协同特性,并采用聚集性评分进一步筛选出任务相关性最强的特征模式,从而实现了任务相关激活区域的高精度检测。实验结果表明,RS-GEm方法能够准确捕捉动作任务中的运动皮层激活区域,在人类连接组数据集上的重测数据中表现出显著的鲁棒性和空间一致性。与GLM方法相比,RS-GEm方法在特征模式提取和任务相关性分析中展现出更高的生物学合理性,同时对噪声的敏感性更低,在构建高精度的任务相关功能图谱方面具有优势。本文提出的RS-GEm方法为任务态fMRI数据分析提供了一种全新的数学框架,不仅在捕捉复杂神经活动非线性特性、解析任务相关功能网络动态变化方面具有重要理论意义,还为研究任务相关脑区功能活动及神经网络的多层次动态特性提供了新的工具支持。此外,该方法在多模态神经影像数据联合分析、神经疾病预测以及复杂任务条件下的脑功能解析中也具有潜在的应用价值。更多还原

【Abstract】 Functional magnetic resonance imaging(fMRI)is an essential tool for investigating neural activities and the dynamic changes in functional brain networks.By measuring blood oxygen level-dependent(BOLD)signals,fMRI captures hemodynamic changes caused by neuronal activity,thereby providing an indirect means of inferring neural activities.Traditional General Linear Model(GLM)approaches rely on linear assumptions,making it challenging to fully capture the nonlinear characteristics of complex neural activities,which presents significant limitations in brain function analysis.As an unsuper-vised dimensionality reduction method,manifold learning can extract low-dimensional nonlinear structures from high-dimensional data,demonstrating unique advantages in representing and analyzing complex neural activities.However,existing manifold learning methods face challenges in processing subject-specific fMRI data due to insufficient utilization of the high-dimensional spatial structure and temporal dependencies of fMRI signals,leading to limited adaptability and feature extraction capabilities.To address these issues,this dissertation proposes a novel framework,including a manifold learning-based feature pattern extraction method and a feature pattern-driven regional synchrony detection approach.The former provides a structured representation,while the latter leverages the extracted feature patterns to achieve high-precision detection of task-related neural activities and detailed analysis of brain functional networks.This dissertation first introduces a method for extracting Brain Graph Eigenmodes based on Laplacian Eigenmaps(BGEm-LE)to detect deep manifold representations of brain neural activities.By preprocessing subject-specific fMRI data and performing data augmentation,combined with low-dimensional embedding analysis using manifold learning methods,this dissertation explores the potential and limitations of Locally Linear Embedding(LLE)and Isometric Mapping(Isomap)in capturing the structure of brain functional networks.The analysis shows that although classical methods can reveal dynamic changes in functional networks under task conditions,they have limitations in eigenmode separation and task-relevant feature extraction.Based on these findings,this dissertation proposes the BGEm-LE method,which constructs subject-specific weighted functional connectivity graphs based on group-level binary brain graphs and applies Laplacian eigen-mode decomposition to extract a set of biologically meaningful brain eigenmodes.Experimental results demonstrate that the BGEm-LE method outperforms LLE and Isomap in functional region segmentation and eigenmode separation,achieving higher resolution and stronger biological relevance.Based on the eigenmodes extracted by BGEm-LE,this dissertation designs a method called Regional Synchronization based on Graph Eigenmodes(RS-GEm)for accurately identifying task-related activation regions to detect brain activity.This method integrates the functional domain information of the extracted eigenmodes with anatomical constraints and a regional synchronization measurement framework to quantify the level of coordinated activity within the functional neighborhood of specific brain regions.Specifically,the RS-GEm method limits the calculation range of functional neighborhoods using geodesic distance to suppress the interference of long-range spurious connections,ensuring the physiological validity of the analysis results.Furthermore,RS-GEm captures the coordinated characteristics of voxels and their functional neighborhoods through synchronization measurements and employs an aggregation scoring mechanism to select the most task-relevant eigenmodes,enabling high-precision detection of task-related activation regions.Experimental results indicate that RS-GEm accurately identifies motor cortex activation regions during motor tasks and exhibits significant robustness and spatial consistency in test-retest data from the Human Connectome Project(HCP)dataset.Compared to the GLM approach,RS-GEm achieves higher biological interpretability in eigenmode extraction and task-related analysis,while being less sensitive to noise,and provides advantages in constructing high-precision task-related functional brain maps.The RS-GEm method proposed in this dissertation provides a novel mathematical framework for task-based fMRI data analysis.It has significant theoretical implications for capturing the nonlinear characteristics of complex neural activities and analyzing the dynamic changes in task-related functional networks.Additionally,it offers a new tool for studying task-related brain functional activities and the multi-level dynamic properties of neural networks.Furthermore,this method holds potential applications in multi-modal neuroimaging data integration,neurological disease prediction,and brain function analysis under complex task conditions.更多还原