【作者】 陈华富；

【导师】 尧德中；

【作者基本信息】 电子科技大学 ， 生物医学工程， 2004， 博士

【摘要】 脑功能磁共振(Functional magnetic resonance imaging fMRI)主要依据血氧水平依赖性(blood oxygenation level dependent,BOLD)对比增强原理进行成像,是目前人们掌握的唯一无侵入、无创伤、可精确定位的研究手段,具有很高的空间分辨率和存在进一步提高时间分辨率的潜力,非常适合神经活动的时空分析和脑的高级功能研究,已受到神经、认知和临床等领域的极大关注。 本文围绕功能磁共振成像(fMRI)在脑科学研究中的应用,对脑功能的神经解剖定位方法、fMRI响应的模型以及fMRI图像的配准等几个fMRI数据分析的主要方面,进行了比较系统性的创新发展;并利用视/听觉fMRI数据验证了所发展的定位方法与动力学响应模型的正确性和有效性。在此基础上,结合临床和认知科学研究中的前沿论题,成功开展了临床癫痫病人的病灶定位分析和脑功能上下左右不对称性的研究。下面分列如下: 1) 在系统分析独立成分分析技术(Independent Component Analysis,ICA)的基本理论和算法的基础上,提出了基于信息极小理论的梯度算法,基于极大似然估计理论的牛顿快速算法和BFGS快速算法,以及结合国际上通用的梯度算法和固定点法优点的组合ICA算法,该组合算法很好地兼顾了ICA算法的稳定性和计算效率。 2) 在ICA算法研究的基础上,深入分析了基于ICA的fMRI数据处理的模型,提出了微空域中的信号与噪声的时域过程相互独立的信号模型,从而建立了一种新的fMRI数据处理方法:邻域独立成分相关法,并从理论、仿真和fMRI数据分析实验几个方面阐明了新方法的合理性和有效性。 3) 针对fMRI的动力学响应中存在不同步和不同响应模式的现象,提出了一种组合空间ICA分离成分进行脑功能定位的新观点,然后结合癫痫活动的特点,建立了一种创新的癫痫活动时空定位方法,并利用实际数据证明了方法的正确性和有效性。 4) 针对fMRI数据的主成分分析,提出了一种新的延时子空间分析方法,能有效地提取fMRI中的弱信号和实现脑功能活动的定位,主成分分析是它的一种特殊情况,从而有效地扩展了fMRI数据的主成分分析方法。更多还原

【Abstract】 Functional magnetic resonance imaging (fMRI) is mainly based on blood oxygenation level dependent (BOLD). It is the most efficient method that can be used to precisely locating brain activities without invasion. With very high spatial resolution and potential high temporal resolution, fMRI is well fit for the spatial and temporal analysis of neural action and the research of advanced brain function. So it is being concerned by many science branches such as neuroscience, cognition and clinic et al and is a hot point in current brain research.Focused on the application of fMRI in brain science, systematic improvements are conducted from neural anatomic localization, fMRI signal model and image registration preprocessing of fMRI. These methods are evaluated with practical visual and auditory fMRI data. Meanwhile, innovative applications are also conducted in a clinical epileptic localization and a study of the brain asymmetry of the left-right and upper-lower visual field. The details are shown as follow:1) Newly proposed is an efficient gradient algorithm of Independent Component Analysis (ICA) based on minimum mutual information, a Newton iterate fast algorithm and BFGS algorithm of ICA based on information maximum likelihood estimation. And a composite ICA algorithm is proposed by combining the merits of the gradient-based algorithm and the fix-point based algorithm, it is of good stability and efficiency.2) Based on the study of ICA algorithm, signal processing models of fMRI data were analyzed. And a signal model based on the independence of temporal courses between signal and noises within a tiny spatial domain were developed, thus a new fMRI data processing: independence component correlation algorithm in a Tiny spatial domain. The new method was evaluated and confirmed by simulation, and real fMRI data.3) Based on the fact that there are asynchronous activations or different response patterns in fMRI process, Proposed is a brain functional localizationmethod, which is an ICA component selection and combination strategy. Applying the method to epileptic activities, a new method of epileptic activities spatial-temporal localization was developed, and its correctness and validity were confirmed by experimental data.4) Based on fMRI data PCA method, proposed is a new delay subspace analysis method, which can pick put weak useful signal and locating brain function activation, PCA is its special instance. So PCA precess method of fMRI data is effectively expanded.5) Based on fMRI data cluster analysis method, proposed is integrated between neighborhood correlation and hierarchical clustering method. Firstly, the neighborhood correlation is implemented to principium imaging, which remove a lot of no activation points, then the hierarchical clustering method is implemented to get the finally imaging image. This method can solve the huge data problem of hierarchical clustering and separate effectively the different patterns of brain activation.6) Besides the above data-driven methods of fMRI signal processing listed in the 2) and 3) above, proposed is a model-driven method, that is a general linear model consisted of the convolution between a new dynamic function and the design matrix. Its validity was confirmed by real visual fMRI data.7) In fMRI signal, there have not only the spatial information of the brain functional activation, but also some more implicit dynamic information of the brain function. Combining Friston’s BOLD microscopic dynamic blood model and Agnes Aubert’s coupling model of brain electrical activity and metabolism, an extended dynamic BOLD model was proposed, which connected brain metabolism with the blood flow blood volume, thus extended the Friston’s model one step ahead in electrophysiological aspect.8) According to the dynamic characteristic of BOLD -fMRI signal, a new model of the fMRI signal was proposed as a convolutions between a Gaussian function and the perfusion function which characterizes the neural response to a stimulus of a neural mass. And an improved Gamma convolution signal model where the convolution is between a Gamma function and the perfusion function更多还原