基于磁共振成像的多变量模式分析方法学与应用研究

发布时间：2018-07-12 12:17

本文选题：多元模式分析 + 多模态分析　；参考：《电子科技大学》2014年博士论文

【摘要】：通过影像数据分析,传统基于组比较的单元分析发现神经精神疾病大脑结构和功能的改变。但是单元分析方法只能在组水平进行推断,导致这些发现对临床诊断的价值非常有限。而且,目前多数神经精神疾病的诊断都依据其临床症状,还没有客观的生物学标记物。因此,如果想让神经影像学的发现更好地应用于临床诊断,就必须提供个体水平的预测。本文主要以磁共振数据为载体,以多元模式分析(Multivariate Pattern Analysis,MVPA)方法学为手段,介绍了结构特征、功能特征,以及结构-结构、功能-功能、功能-结构特征融合在脑模式识别研究中的应用。同时,在不同程度上对MVPA方法进行了改进和创新,将其运用到神经、精神疾病中进行个体水平的诊断并探测这些疾病的病理生理机制。本文内容主要包括5个部分:1.针对不同治疗反应的重度抑郁症患者(Major Depressive Disorder,MDD)的结构磁共振图像(Magnetic Resonance Imaging,MRI)数据,提出Searchlight算法与主成分分析(Principal Component Analysis,PCA)相结合的特征选择方法。从脑结构MRI中提取灰质、白质体积作为特征,使用提出的方法进行特征选择并用支持向量机(Support Vector Machine,SVM)进行分类。实验结果表明提出的MVPA方法优于其它比较流行的方法。采用灰质与白质体积作为特征信息区分不同治疗反应MDD患者的准确率均为82.9%。另外,采用灰质体积特征信息从健康对照中区分难治型、易治型MDD的准确率分别为85.7%和82.4%;采用白质体积作为特征信息从健康对照中区分难治型、易治型MDD的准确率分别为85.7%和91.2%。额、顶、颞、枕叶和小脑一些区域的灰质和白质体积对MDD具有较高的诊断和预后能力。该方法可能为MDD的诊断和预后提供了一个新途径。2.针对社交焦虑障碍(Social Anxiety Disorder,SAD)的静息态功能MRI数据,提出使用大尺度功能脑网络对其建立诊断模型的方法。通过静息态功能MRI数据构建大尺度功能连接网络并将其作为分类特征。然后,采用F分值法进行特征排序并利用SVM进行分类。实验结果表明对SAD患者的正确区分率为82.5%,敏感度为85%,特异度为80%。同时,发现用于区分SAD病人的一致连接主要位于几个静息态网络内部或者之间的连接,包括:默认网络、视觉网络、感觉运动网络、情感网络以及小脑区。此外,右侧眶额皮层在分类过程中占了最高的权重。这些发现为确定SAD潜在的生物学标记物提供了一定的依据。3.针对传统基于LASSO特征选择法的局限性,提出高阶图匹配的特征选择方法,并用老年痴呆症(Alzheimer's Disease,AD)神经影像学(Alzheimer's Disease Neuroimaging Initiative,ADNI)数据进行验证。基于LASSO的特征选择法对每个样本的目标向量进行独立的估计而没有考虑与其它样本的联系,从而忽略了训练集目标向量之间的几何关系。同时,预测向量与目标向量应该有相似的几何关系。将这个问题看作预测图与目标图之间的图匹配问题,通过提出二元关系正则项和三元关系正则项解决了LASSO特征选择法的不足。本文采用灰质体积和皮层厚度作为分类特征,由高阶图匹配方法对两种特征分别进行特征选择并用多核学习法进行特征融合。该方法对AD和轻度认知障碍(Mild Cognitive Impairment,MCI)分类分别得到了92.17%和81.57%的准确率,优于基于LASSO的特征选择法,这验证了方法的有效性。4.针对创伤后应激障碍(Post-traumatic Stress Disorder,PTSD)的静息态功能MRI数据,提出融合多水平特征对其进行分类的方法。从静息态功能MRI中提取出3个水平(区域内,区域间和全脑)的特征,采用t检验与SVM递归特征消除(Recursive Feature Elimination,RFE)相结合的方法进行特征选择,并用多核SVM融合多水平功能特征进行分类。实验结果表明每个水平的特征都能成功的区分PTSD病人,通过多水平特征的融合可以进一步提高分类的性能。所提出的模型对PTSD分类得到的准确率为92.5%,比只使用2个水平特征和1个水平特征的准确率分别至少高5%和17.5%。而且,发现边缘系统和前额叶皮层为分类提供了最具有区分力的特征。该研究可能为改善PTSD的临床诊断提供了一个补充的方法。5.针对以往多模态数据分类问题中特征选择的局限性,提出约束模态间关系的多模态多任务特征选择方法,并使用ADNI数据进行验证。传统的多模态分类问题中的特征选择法往往在每个模态内部单独进行,并没有考虑到不同模态之间特征选择的关系。因此,提出将每个模态中进行的特征选择作为一个任务,在特征选择时对模态间的关系进行约束,并保持模态内部选择特征的稀疏性。在特征方面,从正电子发射断层成像(Positron Emission Tomography,PET)中提取出区域平均代谢强度,结构MRI中提取出区域平均灰质体积作为分类特征。由提出的方法进行特征选择并用多核SVM进行多模态特征融合。实验结果表明,对AD的分类准确率达到了94.37%,MCI的分类准确率达到了78.80%,MCI转化组和非转化组的分类准确率达到了67.83%。这些结果显著优于传统的特征选择方法,这证实了所提出方法的优越性。
[Abstract]:Through the analysis of the image data, the traditional unit analysis based on the group comparison found the changes in the brain structure and function of the neuropsychiatric disorders. But the method of unit analysis can only be inferred at the level of the group, resulting in the very limited value of these findings to the clinical diagnosis. And the diagnosis of most psychic diseases is based on its clinical symptoms, There is no objective biological marker. Therefore, if we want to make the findings of neuroimaging better applied to clinical diagnosis, it is necessary to provide the prediction of individual level. This paper mainly uses the magnetic resonance data as the carrier and the Multivariate Pattern Analysis (MVPA) methodology as a means to introduce the structural features and functional characteristics. And the application of structure structure, structure, function, function, functional structure feature fusion in brain pattern recognition research. At the same time, the MVPA method is improved and innovating to different degrees. It is applied to the individual level diagnosis and detection of the pathophysiological mechanism of these diseases in the nerve and mental diseases. The main contents of this paper include 5 parts. 1. in view of the structural magnetic resonance imaging (Magnetic Resonance Imaging, MRI) data of Major Depressive Disorder (MDD) in patients with different therapeutic responses (Magnetic Resonance Imaging, MRI), a feature selection method combining Searchlight algorithm with principal component analysis (Principal Component Analysis) is proposed. As a feature, the proposed method is used for feature selection and classification with Support Vector Machine (SVM). The experimental results show that the proposed MVPA method is superior to other popular methods. The accuracy of using gray matter and white matter volume as the characteristic information to distinguish different treatment responses to MDD patients is 82.9%.. The accuracy rates of MDD were 85.7% and 82.4%, respectively, with the gray matter volume information from the healthy control area. The white matter volume was used as the characteristic information from the healthy control middle area. The accuracy rate of the MDD was 85.7% and 91.2%. respectively. The gray matter and white matter volume in the top, temporal, occipital and cerebellum areas were MDD It has a high diagnostic and prognostic ability. This method may provide a new approach for the diagnosis and prognosis of MDD,.2. for the resting state functional MRI data of social anxiety disorder (Social Anxiety Disorder, SAD), and proposes a recipe for the establishment of a diagnostic model using a large scale functional brain network. The construction of large scale work through the resting state function MRI data. We can connect the network and use it as a classification feature. Then, the F segmentation method is used to sort the features and use SVM to classify them. The experimental results show that the correct discrimination rate for SAD patients is 82.5%, the sensitivity is 85%, the specificity is 80%., and the consistent connection used to distinguish the SAD patients is mainly located within or between several resting networks. The connection, including the default network, the visual network, the sensory network, the emotional network and the cerebellum. In addition, the right orbital frontal cortex accounts for the highest weight in the classification process. These findings provide a certain basis for determining the potential biological markers of SAD based on the limitations of the.3. based on the LASSO feature selection method. The matching feature selection method is validated with Alzheimer's Disease (AD) neuroimaging (AD) neuroimaging (Alzheimer's Disease Neuroimaging Initiative, ADNI) data. The feature selection method based on LASSO is used to estimate the target vectors of each sample independently without considering the connection with other samples, thus neglecting the training. At the same time, the geometric relationship between the target vectors and the prediction vector should be similar to the target vector. The problem is regarded as a graph matching problem between the prediction graph and the target graph. The deficiency of the LASSO feature selection method is solved by putting forward the regular term of the two element relation and the regular term of the three element relation. This paper uses the gray matter volume and the cortex. As a classification feature, the two features are selected by the high order graph matching method and the multi kernel learning method is used for feature fusion. The accuracy of the method for AD and Mild Cognitive Impairment (MCI) classification is 92.17% and 81.57% respectively, which is superior to the LASSO based feature selection method, which verifies the method. The effectiveness of.4. is based on the resting state function MRI data of post traumatic stress disorder (Post-traumatic Stress Disorder, PTSD), and proposes a method of classifying it with multi level features. 3 levels (intra, interregional and whole brain) are extracted from the resting state function MRI, and t test and SVM recursion feature elimination (Recursive Featur) are used. E Elimination, RFE) combines the feature selection method and classifies the multilevel function features with multi core SVM fusion. The experimental results show that the characteristics of each level can successfully distinguish the PTSD patients. The classification performance can be further improved by the fusion of multilevel features. The accuracy rate of the proposed model for PTSD classification can be further improved. For 92.5%, the accuracy rate of only 2 horizontal and 1 horizontal features is at least 5% and 17.5%., respectively, and it is found that the edge and prefrontal cortex provide the most regional characteristics for classification. This study may provide a supplementary method for improving the clinical diagnosis of PTSD,.5. for the problem of the previous multi-modal data classification problem. In the limitation of feature selection, a multi modal and multi task feature selection method that constrains the inter modal relationship is proposed, and the ADNI data is used to verify. The feature selection method in the traditional multi-modal classification problem is often carried out separately in each mode, and does not take into account the relationship between different modes. The feature selection in the mode is used as a task to restrict the relationship between the modes and maintain the sparsity of the selection characteristics in the mode. In the characteristic aspect, the average metabolic intensity of the region is extracted from the Positron Emission Tomography (PET), and the regional average is extracted from the structure MRI. The gray matter volume is classified as the classification feature. The method is selected by the proposed method and multimodal SVM is used to fuse the multi-modal features. The experimental results show that the classification accuracy of AD is 94.37%, the classification accuracy of MCI reaches 78.80%, and the classification accuracy of the MCI transformation group and the non transformation group reaches 67.83%. these results are significantly better than the traditional ones. The method of feature selection confirms the superiority of the proposed method.
【学位授予单位】：电子科技大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：R445.2;R741

【共引文献】