CASIA OpenIR  > 毕业生  > 博士学位论文
功能磁共振成像数据的稀疏表示与字典学习研究
王鑫
学位类型工学博士
导师张文生
2017-05-27
学位授予单位中国科学院研究生院
学位授予地点北京
关键词功能磁共振成像 稀疏表示 字典学习 疾病分类 功能脑网络
摘要人类大脑是自然界中最复杂的信息处理系统之一,探索脑功能的奥秘是脑科学领域的一个研究热点。近年来,功能磁共振成像( functional Magnetic Resonance ImagingfMRI)技术凭借非侵入式、较高的时间和空间分辨率等优点,成为一种理解脑功能的有力工具。探索 fMRI 数据的分析方法,对于理解大脑的信息处理机制,分析神经和精神疾病的脑功能异常具有重要意义。
在设计
fMRI 数据的分析方法时,结合大脑的神经元激活特点可以得到更符合 fMRI 数据分布的结果。神经科学研究表明,大脑的神经元激活具有稀疏响应的特点。因此,稀疏表示及字典学习方法成为医学影像数据分析的重要工具,在医学图像处理领域中得到了广泛的关注和应用。本文利用人脑神经元激活稀疏性的生理学发现,探索适用于 fMRI 数据分析的稀疏表示及字典学习方法,以便于帮助我们更好地理解大脑的信息处理机制,为神经和精神疾病的诊断与治疗提供理论基础。
本文主要工作与创新点如下:

1. 提出一种结合稀疏低秩模型和基于图的特征的疾病分类方法,用于解决传统方法在构建功能脑网络过程中只考虑两两脑区的功能连接以及从功能脑网络中提取单一局部特征的问题。该方法采用稀疏低秩模型构建功能脑网络,在估计两个脑区的功能连接时考虑其他脑区的作用。同时,模型中的稀疏和低秩约束可以令构建的功能脑网络有更强的稀疏性和模块化结构,使其更符合功能脑网络固有的结构特点。此外,该方法结合多种基于图的特征,从多角
度挖掘功能脑网络中的有效信息,有利于提升分类效果。在抑郁症
fMRI 数据集上的实验结果表明,该方法的分类效果优于对比方法。
2. 提出一种多任务融合最小绝对值收缩和选择算子( Least Absolute Shrinkage and Selection OperationLasso)模型的动态功能脑网络构建方法,用于解决静态功能脑网络的构建模型忽略功能连接随时间变化的动态信息的问题。该方法采用多任务融合 Lasso 模型构建动态功能脑网络,保留功能脑网络的稀疏性与子序列功能连接的时间平滑性,有利于刻画脑区之间功能连接的稀疏性与相邻 fMRI 子序列之间的相似性特点。在 ADNI 的公开 fMRI 数据集上的分类效果验证了多任务融合 Lasso 模型构建动态功能脑网络的有效性。
3. 提出一种基于加权判别式字典学习的疾病分类方法,用于解决传统的字典学习方法忽略了样本与字典之间重要关系信息的问题。该方法通过引入自适应的加权机制,利用样本与字典之间的相似性对不同样本的表示残差进行差异性惩罚,从而提高字典的判别性。此外,该方法用 2-范数正则化约束表示系数,以尽量避免小样本引起的过拟合现象。在抑郁症 fMRI 数据集和ADHD-200 的公开 fMRI 数据集上的实验结果表明,与对比方法相比,该方法有更快的算法速度和更好的分类效果。
其他摘要The human brain is one of the most complex systems of information processing in nature. Exploring the mystery of brain function is a research hotspot in the field of brain science. In recent years, functional magnetic resonance imaging (fMRI), given its non-invasiveness and high temporal/spatial resolution, has emerged as a powerful tool in understanding the brain function. Exploring the data analysis methods based on fMRI is very important for understanding the brain information processing mechanisms and the functional abnormalities of neurological or psychiatric disorders.
Methods which consider the characteristics of brain neuronal activation can achieve results that are in line with the fMRI data distribution. Many neuroscience studies indicate that the brain neuronal activation is a sparse response. Sparse representation and dictionary learning have become an important tool for the analysis of medical image data. Sparse representation and dictionary learning have been widely used in the field of medical image processing. In this dissertation, we explore the sparse representation and dictionary learning methods for fMRI data, based on the sparsity of brain neuronal activation in the neuroscience findings. The aim of our study is to achieve a better understanding of the brain information processing mechanisms and to provide the oretical basis for the diagnosis and treatment of neurological and psychiatric disorders.
The main contributions of this dissertation can be summarized as follows:
1. We present a disease classification method which combines sparse low-rank model and graph-based features. This method is proposed to overcome the limitations lying in traditional methods, in which only pairwise relationship between brain regions is captured and single local feature is extracted. The functional brain network (FBN) is constructed by a sparse low-rank model, which considers the relationship between two brain regions given all the other brain regions. Moreover, the sparse and low-rank constraint in the model can result in a more sparse and clearer modular structure in the FBN, which conforms to the inherent structures of FBN. In addition, eight graph-based
features are extracted from FBN to exploit effective information of FBN from different aspects, which can improve the classification performance. Tested on a fMRI dataset of depression, our method performs better than the compared methods.
2. We propose a multi-task fused least absolute shrinkage and selection operation
(Lasso) method to construct dynamic FBN, which can mine the dynamic information of functional connectivity. The multi-task fused Lasso method can preserve the sparsity and temporal smoothness of dynamic FBN, which can measure the sparsity of functional connectivity and similarity of adjacent fMRI sub-series. The classification performance on a public fMRI dataset from ADNI further verify the effectiveness of our method for constructing dynamic FBN.
3. We put forward a weighted discriminative dictionary learning method for disease classification. This method is presented to address the issue that traditional dictionary methods often ignore the valuable relationship between samples and dictionary. An adaptive weighting scheme is introduced to exploit the underlying relationship between samples and dictionary. The introduction of weighting scheme discriminately penalizes representation errors of different samples, which makes the dictionary more discriminative. Meanwhile, the 
2-norm regularization constraint on coding coefficients is adopted to avoid overfitting caused by small sample size. The experiments are performed on a fMRI dataset of depression and a public fMRI dataset from ADHD-200. Results suggest that the proposed method provides faster and better classification performance than the compared methods.


语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/14843
专题毕业生_博士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
王鑫. 功能磁共振成像数据的稀疏表示与字典学习研究[D]. 北京. 中国科学院研究生院,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
功能磁共振成像数据的稀疏表示与字典学习研(22807KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[王鑫]的文章
百度学术
百度学术中相似的文章
[王鑫]的文章
必应学术
必应学术中相似的文章
[王鑫]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。