基于静息态功能磁共振成像的自闭症预测研究

发布时间：2018-06-13 05:58

本文选题：自闭症预测 + 静息态功能连接　；参考：《北京交通大学》2017年硕士论文

【摘要】：自闭症(Autism spectrum disorder,ASD)是一种神经发育障碍疾病,发病率高达1%,给社会及患者的家庭带来了沉重负担。目前主要以行为量表对ASD进行诊断,具有一定的主观性。静息态功能连接(Resting-state functional connectivity,RSFC)反映了大脑在静息态下不同脑区神经活动模式之间的时间相关性,基于RSFC探索能够识别ASD的生物标记对于ASD的客观辅助诊断和理解其神经机制具有重要意义。本文基于RSFC从静态和动态两方面进行了 ASD预测研究:一是认为RSFC在整个扫描过程中是静止的,利用Lasso和elastic net两种方法对RSFC的特征选择进行了深入研究;二是假设RSFC是随时间动态变化的,基于动态功能连接方法对ASD预测进行了初步探索,具体工作内容如下:(1)针对大部分方法不能有效地选出具有识别力的RSFC,本文提出利用Lasso选择有识别力的RSFC。首先计算Pearson相关捕捉到大脑的正负相关RSFC,然后进行阈值化保留同步化程度较高的正相关RSFC,进一步利用嵌入式特征选择方法Lasso去除冗余的RSFC只保留最有识别力的特征子集,最后基于SVM分类得到81.52%的分类准确率,同时找出了 19个具有显著识别力的RSFC。(2)针对Lasso方法无法处理具有组效应的变量选择问题,进一步提出基于elastic net的多级特征选择方法进行ASD预测研究。本文依次利用阈值化、t检验和elastic net逐步筛选出差异越来越显著的RSFC特征子集。t检验能初步筛选出与临床症状显著相关的RSFC;Elastic net能发挥处理组效应变量选择的优势对RSFC作进一步筛选。最终ASD预测的准确率达到84.78%,进一步提升了预测性能,并确定了 22个有显著差异的RSFC。(3)针对大部分研究主要基于静态功能连接方法进行ASD预测,而动态功能连接比静态蕴含的信息更为丰富,本文基于动态功能连接分别提取了高阶功能连接特征和动态网络拓扑特征(节点连接度)进行ASD预测。高阶功能连接能捕获各脑区低阶功能连接之间的时间相关信息,最终得到81.52%的预测准确率;动态网络拓扑方法能提取拓扑结构随时间的动态变化信息,预测性能还不够理想。在提出的两种静态功能连接方法中,elastic net方法比Lasso获得了更好的预测性能,有助于寻找到与ASD有关的生物标记以辅助医生进行临床诊断。高阶功能连接方法也获得了不错的预测性能,虽不及静态方法,但给我们提供了另一个思路去寻找生物标记,有利于发现ASD的隐含神经机制。
[Abstract]:Autistic spectrum disorder (ASD) is a neurodevelopmental disorder with a high incidence of 1, which brings a heavy burden to the society and the families of the patients. At present, the diagnosis of ASD is mainly based on behavior scale, which is subjective. Resting state functional connectivity (RSFCs) reflects the temporal correlation of neural activity patterns in different regions of the brain under resting state. Exploring biomarkers that can identify ASD based on RSFC is of great significance for objective diagnosis and understanding of the neural mechanism of ASD. In this paper, the static and dynamic aspects of ASD prediction are studied based on RSFC. First, it is considered that RSFC is stationary in the whole scanning process, and the feature selection of RSFC is studied deeply by using Lasso and elastic net methods. On the other hand, assuming that RSFC changes dynamically with time, the prediction of ASD is preliminarily explored based on dynamic functional connection method. The main work of this paper is as follows: (1) aiming at the fact that most of the methods can not effectively select the RSFCs with recognition power, this paper proposes to use Lasso to select the discriminative RSFCs. Firstly, Pearson correlation was calculated to capture positive and negative correlation RSFCs in the brain, then threshold retention synchronization of positive correlation RSFCs was performed. Further, the embedded feature selection method Lasso was used to remove redundant RSFC, which only retained the most recognizable subset of features. Finally, 81.52% of the classification accuracy is obtained based on SVM classification. At the same time, 19 RSFC.2s with significant recognition power are found. For Lasso method, it can not deal with variable selection problem with group effect. Furthermore, a multilevel feature selection method based on elastic net is proposed. In this paper, we use the threshold t test and elastic net step by step to screen out the characteristic subset. T test, which is more and more significant. The result shows that RSFCU Elastic net, which is significantly related to clinical symptoms, can play an important role in the selection of effect variables of the treatment group. RSFC was further screened. Finally, the accuracy rate of ASD prediction reaches 84.78, which further improves the prediction performance, and determines 22 RSFC.K3, which have significant differences. In view of the majority of researches, ASD prediction is mainly based on static functional connection method. The dynamic functional connection is more abundant than the static information. Based on the dynamic functional connection, the high-order functional connection feature and the dynamic network topology feature (node connectivity degree) are extracted to predict the ASD. High-order functional connections can capture the time-dependent information among the low-order functional connections in various brain regions, and finally obtain 81.52% prediction accuracy. The dynamic network topology method can extract the dynamic changes of topology structure with time, and the prediction performance is not good enough. In the two static functional connection methods proposed, the modified net method has better predictive performance than Lasso, which is helpful to find the biomarkers associated with ASD-related biomarkers to assist doctors in clinical diagnosis. High-order functional join method also has good predictive performance. Although it is not as good as static method, it provides us with another way to find biomarkers, which is helpful to discover the implicit neural mechanism of ASD.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R445.2;R749.94

【参考文献】