毕业生知识库

精神分裂症是一种复杂的精神疾病。探究精神分类症发病机理不仅有助于疾病本身的诊断和治疗，也会促进人对人脑这一复杂并且至关重要的器官的理解。功能磁共振成像技术以及基于静息态功能成像技术的脑网络方法逐渐成为探究这一疾病的病理生理机制的重要手段。现有的研究支持精神分裂症是一种失连接的精神疾病，主要体现在脑网络结构之间的结构连接或者功能连接的异常。然而囿于精神分裂症本身的复杂性，现有观测技术的局限，以及已有研究样本数的局限，过往研究结果给我们的结论有限而模糊。而针对精神分裂症的不同维度症状进行自下而上、各个击破的研究，不失为理解这种复杂疾病的有效路径。此外，随着该领域的推进，研究者们开始运用机器学习技术等先进的数据分析手段分析影像数据，有助于直接探索自动诊断精神疾病的方法。本文旨在基于静息态功能连接的脑网络分析方法，探究精神分裂症的脑网络受损特征，特别是针对工作记忆核心网络，并试图利用机器学习技术来获得自动诊断方法。本文的具体研究内容和创新之处如下：  
1. 基于激活共变模式的分区方法。选择合适的脑区结构作为脑网络的节点是脑网络分析的基础。一般而言，我们强调每个节点是功能相对独立的脑区。因此，研究如何对人脑进行分区是脑网络技术的课题之一。在本项工作中，我们提出了一种利用与感兴趣区域相关的任务实验做荟萃分析进行分区的方法。通过估计每个体素在各个实验中的激活值作为它们的激活模式，进而计算两两体素的相似度来聚类得到划分的子区。我们比较了新的方法和细胞构筑以及其他基于文献坐标方法的分区结果，证明了该方法的有效性和优势。此外，利用文献中的实验任务信息，通过比较每个子区和整个感兴趣区域激活点对应实验的分布，可以推断每个子区的功能画像。  
2. 考察了精神分裂症工作记忆“核心网络”内部的功能连接特征。基于现有荟萃分析定义的工作记忆“核心网络”，主要位于额叶、顶叶、脑岛等区域。我们比较了这些节点之间功能连接在精神分裂症患者和正常人两组间的差异。研究发现，精神分裂症患者在四条连接中表现出功能整合能力减弱（连接降低），这些功能连接的受损可能意味着患者在执行工作记忆任务时对无关信息的抑制能力以及注意力的分配机制受到损害。得益于多中心大样本数据，我们在统计分析中使用了多中心荟萃分析的方法，提高了统计效力。  
3. 基于多任务学习方法对多中心大样本精神分裂症数据作特征选择和分类。以脑网络为特征对精神分类症样本分类是一个高维小样本的分类问题。特征选择这一步是至关重要的一个步骤。考虑到由于不同来源的精神分裂症脑影像数据的异质性，导致分别在每个数据集上单独作特征选择的不一致，我们采用了多任务学习方法，即通过L12范数约束分类器在不同数据集上所选择的特征的一致性，使得所选特征是在所有数据集上对分类都有贡献的特征。我们发现几条功能连接对分类任务有一致稳定的贡献，这些功能连接可能反映精神分裂症最核心的病理特征。同时我们发现，为了保证较高分类精度，分类器必须纳入绝大部分功能连接特征。这些特征广泛分布于全脑，因此我们推测精神分裂症是一种弥散性的失连神经生理精神疾病。

Schizophrenia is a complex mental illness. The endeavor to exploring the pathogenesis of mental disorders not only contributes to the diagnosis and treatment of the disease itself, but also promotes the understanding of human brain, as the most complex and vital organ of human. Brain network analysis based on Functional magnetic resonance imaging (FMRI) has become an important method to explore the pathophysiology of this disease.

The existing studies support schizophrenia as a mental disorder of improper functional integration, manifested by abnomral structural or functional connections between brain network structures. However, due to the complexity of schizophrenia itself, the limitations of the existing observation technology and small the sample size, the results of previous studies give us limited and vague conclusions. A bottom up and point to surface strategy by forcusing on each dimension of the schizophrenia symptoms may be an effective path to explore such a disease. In addition, with the advance of the field, the researchers began to use machine learning technology and other advanced data analysis tools to to directly solve the problem of automatic diagnosis of mental illness. The aims of this dissertation are to investigate the imparied resting-state functional connections in schizophrenia, especially for the working memory core network, and try to use machine learning techniques to explore the ways of automatic diagnosis of schizophrenia. The main contents and contributions of this dissertation are as follows:

1. Meta-analytic activation modeling-based parcellation (MAMP). Choosing the appropriate brain structure as the node is important in brain network analysis.In general,we emphasize that each node is a functionally independent brain region. Therefore,the study of how to partition the human brain to such regions is one of the topics of brain network technology. In this work, we propose a new parcellation scheme that uses “modeled activation” pattern across the experiments related with the region of interest (ROI). By estimating the activation value of each voxel in each experiment as its activation mode and calculating the similarity of the two voxels, the clustered sub-regions are obtained. We compare the results with cytoarchitecture results and MACM results, and demonstrate the effectiveness and advantage of this method. In addition, using the experimental task information in the meta data, we can deduce the functional portraits of each subregion by comparing the distributions of the activation points corresponding to each sub-region and the whole region of interest.

2. The functional connectivity working memory “core network” in schizophrenia was investigated. Based on the existing meta-analysis de?ned working memory “core network” that mainly located in the frontal lobe, parietal lobe, and insular, we calculated functional connectivities between these nodes and compaired between the two groups of schizophrenia patients and normal. The study found that patients with schizophrenia showed reduced functional integration (decreased connectivity) in four connections. The impairment of these functional connections may imply that the patient’s ability to suppress irrelevant information in the execution of a working memory task and the allocation of attention are compromised. Oweing to the multicentre large sample data, we were able to use a multicenter meta-analysis stratergy in statistical analysis to improve the statistical power.

3. A multi-task learning method was applied in feature selection and classification on multicenter large sample schizophrenia dataset. It is a high-dimensional and small-sample classfication problem when we use functional connectivitiy features to classify the patients and the normal. Therefore, feature selection is a critical step. Due to the heterogeneity of schizophrenia patients and multicenter imaging, feature selection on each dataset separately result in inconsistency across data sites.

To address this issue, we adopt the multi-task learning method, which adopts L12 norm to constrain the feature selection of different datasets, so that the selected feature contributes to the classification on all data sets. We found that several functional connections have consistent contributions to classification tasks, which may reflect the core pathological features of the psychiatric class. Additionally, we found that, in order to ensure high classification accuracy, the classifier must incorporate most of the functional connection features. These features are widely distributed in the whole brain, so we speculate that schizophrenia is a diffuse neurophysiological psychiatric disorder.