MRI图像分析中的稀疏特征学习方法研究

发布时间：2018-07-07 14:36

本文选题：功能磁共振成像 + 结构磁共振成像　；参考：《北京理工大学》2015年博士论文

【摘要】：近年来,MRI图像分析方法被越来越多地应用于大脑结构和功能研究以及神经性疾病的计算机辅助诊断中。另一方面,随着人工智能的发展,机器学习技术,尤其是其中的稀疏特征学习也被越来越多地引入MRI图像分析中,在分类和预测建模方面扮演着重要角色。因此,研究新的MRI图像分析方法,是更好地深度挖掘MRI图像信息,进而促进脑科学研究及计算机辅助诊断技术发展的关键。然而,MRI图像分析常常面临小样本、高特征维度的问题,由此导致的过拟合、噪声特征和冗余特征严重降低了模型性能。稀疏特征学习方法能够很好地解决上述难点,并已经成功地应用于信号处理、模式识别和计算机视觉领域。本论文致力于研究新的适用于MRI图像分析的稀疏特征学习方法,通过设计新的代价函数约束项来挑选拥有最佳分辨性能的图像特征,从而提升模型分类和预测性能。本研究涉及的稀疏特征学习方法既包括稀疏单任务学习,如稀疏贝叶斯学习和基于1L范数的稀疏学习,也包括稀疏多任务学习,如组Lasso,Dirty model和稀疏组Lasso。通过对上述方法进行不同程度的创新,将其应用于认知神经科学和神经疾病诊断研究中,取得了良好的效果。本文的工作及创新之处主要包括以下5个部分:1.建立了一个基于多体素模式分析的学习模型,在初级视皮层上对空间视觉刺激进行解码研究,解决了单体素分析方法忽视了体素之间的相关信息这一缺点。进一步,建立了一种多分类的稀疏贝叶斯学习模型,将特征选择与视觉解码结合起来。该模型能够在选择最相关特征的同时利用挑选出的特征进行视觉解码,具有很好的整合性。实验结果表明,该方法从2000个初级视皮层体素中挑选出9个最相关体素,使用挑选出的体素进行解码,分类精度达到91.6%。同时,将挑选出的9个体素映射回原始脑空间,从另一个角度验证了初级视皮层具有视网膜映射特性。该方法为基于功能MRI的视觉研究提供了一种新的途径。2.首次提出一种基于结构MRI图像和1L范数稀疏特征学习的烟雾病诊断方法,解决了传统数字造影方法有损、技术复杂及代价高昂的缺点,使得将烟雾病诊断作为常规检查成为可能。具体来说,该方法首先提取结构MRI图像的皮层厚度特征,每幅图像得到约2万个特征。然后建立了三种基于1L范数的稀疏特征学习模型,包括Lasso,弹性网和L1-logistic回归,通过特征选择实现特征约简,最后用挑选出的特征训练支持向量机分类器。实验结果表明,提出的诊断方法取得了较好的诊断精度(分类精度),其中基于弹性网特征学习的方法取得了最高的诊断精度,达到82.36%,对应ROC曲线下的面积0.833,显著优于未经过特征选择直接使用支持向量机对所有提取特征分类的结果(分类精度71.72%,对应ROC曲线下面积0.787)。3.利用儿童(6岁到15岁)的结构MRI图像,建立了基于多核支持向量回归的智商估计模型,并在建模过程中提出一种改进Dirty model多任务特征学习方法用于特征选择,较好地实现了对儿童的智商估计。具体来说,首先提取儿童结构MRI图像的灰质/白质特征,将对灰质/白质的特征选择分别看做一个学习任务,利用提出的改进Dirty model选择与智商相关的灰质/白质特征。分别计算挑选出的灰质/白质特征的核函数,送入多核支持向量回归模型中进行智商估计。实验结果显示,提出的方法估计的智商分数与儿童真实智商分数的相关性为0.718,与真实智商分数之间的均方根误差为8.695。该智商估计模型的意义在于,为今后预测婴幼儿智商,从而根据其预测智商定制合理的早期学习计划提供了方法学基础。4.提出一种基于分层模型和改进Dirty model的多被试解码方法,较好地解决了传统基于功能MRI解码方法需要对每个被试单独建模的缺点。传统的解码方法以体素作为特征,但不同被试之间对于相同实验刺激的体素激活模式差异较大,难以对所有被试建立一个统一的模型。为解决这一问题,提出一种分层模型,以体素作为低层特征,学习出更加鲁棒的高层特征,然后利用提出的改进Dirty model对高层特征进行选择,并使用选择出的高层特征,对所有被试建立一个统一的解码模型。将提出的多被试解码方法用于2D/3D视觉刺激的功能MRI数据进行验证,分类精度达到89.4%,显著高于直接使用体素作为特征的方法(73.4%)。该方法为研究不同被试间的神经特性提供了方法学上的支持。5.提出一种基于结构MRI图像和稀疏特征学习的自闭症诊断方法,为解决传统自闭症诊断方法以行为学评分为标准,而实际中很难用一种行为学评分去诊断该疾病这一问题提供了一种新的方法学途径,并在建模过程中提出一种全新的Canonical图匹配稀疏组Lasso多任务特征学习算法用于特征选择。具体地说,首先提取结构MRI图像的灰质/白质特征,通过典型相关分析将原始灰质/白质特征映射到一个新的Canonical空间。将对类标签以及SRS_TOTAL行为学评分的Canonical特征选择分别看做一个学习任务,利用提出的方法选择与自闭症分类最相关的Canonical特征,然后用支持向量机进行分类。实验结果显示,提出的方法诊断精度为75.4%,ROC曲线下面积为0.804,显著优于基于原始灰质/白质特征的最新方法,也优于现有的其它基于Canonical特征的方法。
[Abstract]:In recent years, MRI image analysis methods have been used more and more in the study of brain structure and function and computer aided diagnosis of neuropathic diseases. On the other hand, with the development of artificial intelligence, machine learning technology, especially the sparse feature learning among them, has also been introduced into MRI image analysis more and more, in classification and prediction Modeling plays an important role. Therefore, the study of the new MRI image analysis method is the key to the better depth mining of MRI image information, thus promoting the development of brain science research and computer aided diagnosis technology. However, MRI image analysis often faces small samples, high characteristic dimensions, resulting in overfitting, noise characteristics and redundancy. The redundant features seriously reduce the performance of the model. The sparse feature learning method can solve the above difficulties well, and has been successfully applied to the field of signal processing, pattern recognition and computer vision. This paper is devoted to the study of the new sparse feature learning method for MRI image analysis and the design of new cost function constraints. The sparse feature learning methods involved in this study include sparse single task learning, such as sparse Bayesian learning and sparse learning based on 1L norm, and sparse multitask learning, such as group Lasso, Dirty model, and sparse group Lasso. through pairs. This method has been applied to cognitive neuroscience and neural disease diagnosis research and has achieved good results. The work and innovation of this paper mainly include the following 5 parts: 1. a learning model based on the model analysis of the multibody element is established to decode the spatial visual stimuli on the primary visual cortex. In addition, a multi classification sparse Bayesian learning model is established, which combines feature selection with visual decoding. The model can be used to select the most relevant features and use the selected features for visual decoding, which is very good. The experimental results show that the method selects 9 most relevant voxels from 2000 primary visual cortex voxels, decodes the selected voxels, and the classification accuracy reaches 91.6%.. The selected 9 individual elements are mapped back to the original brain space and the retina mapping characteristics of the primary visual cortex are verified from another angle. The method provides a new way for visual research based on functional MRI..2. first proposes a method of diagnosis of moyamoya disease based on structural MRI image and 1L norm sparse feature learning. It solves the disadvantages of traditional digital contrast method, which is complicated and costly. It makes it possible to make the diagnosis of smoke fog as a routine examination. For example, this method first extracts the cortical thickness characteristics of the structure MRI image, and each image gets about 20 thousand features. Then three sparse feature learning models based on 1L norm are established, including Lasso, elastic network and L1-logistic regression. Feature reduction is implemented by feature selection. Finally, the selected feature is used to train the support vector machine classifier. The experimental results show that the proposed method has achieved good diagnostic accuracy (classification accuracy), and the method based on the feature learning of the elastic network has achieved the highest diagnostic accuracy, reached 82.36%, corresponding to the area of 0.833 under the ROC curve, and is significantly better than the node without the feature selection using the support vector machine to classify all the extracted features. (classification accuracy 71.72%, corresponding to the area under the ROC curve 0.787).3. using the structure MRI images of children (6 to 15 years old), an IQ estimation model based on multi kernel support vector regression is established. In the modeling process, an improved Dirty model multi task feature learning method is proposed for feature selection, and the IQ estimation of children is better realized. Specifically, the gray matter / white matter features of the children's structure MRI images are first extracted and the selection of gray matter / white matter features is considered as a learning task respectively. The gray matter / white matter characteristics associated with the IQ are selected by the proposed improved Dirty model. The kernel functions of the selected gray matter / white matter are calculated and sent to the multicore support vector back. The results show that the correlation between the estimated IQ score and the true IQ score of the proposed method is 0.718, and the root mean square error between the true IQ score and the true IQ score is 8.695., the IQ estimation model is for the future prediction of infant IQ, so that it is reasonable to customize the IQ according to its IQ prediction. The early learning plan provides a methodological basis for.4. to propose a multi decode method based on a hierarchical model and an improved Dirty model, which is a good solution to the shortcomings of the traditional functional MRI decoding method that needs to be modeled individually for each subject. The traditional decoding method is characterized by voxels, but different subjects have the same experimental stings between the different subjects. In order to solve this problem, a hierarchical model is proposed to learn more robust high level features, and then use the proposed improved Dirty model to select the high level features and use the selected high-level features. A unified decoding model is established for all subjects. The proposed multi test decoding method is used to verify the functional MRI data of 2D/3D visual stimuli. The classification accuracy is 89.4%, which is significantly higher than that of the direct use of voxel (73.4%). This method provides a methodological support for the study of the neural characteristics between different subjects. An autism diagnosis method based on structural MRI image and sparse feature learning is proposed to solve the traditional autism diagnosis method, which is based on behavioral score. In practice, it is difficult to use a kind of behavioral score to diagnose the disease. A new way is provided in the process of modeling, and a new Cano is proposed in the modeling process. Nical graph matching sparse group Lasso multi task feature learning algorithm is used for feature selection. Specifically, first extract gray / white texture features of structural MRI images, map the original gray / white matter features to a new Canonical space through canonical correlation analysis, and select the Canonical features of class labels and SRS_TOTAL behavioral scores. Do not look at a learning task, use the proposed method to select the most relevant Canonical features with the autism classification, and then use the support vector machine to classify them. The experimental results show that the proposed method has a diagnostic accuracy of 75.4% and the area under the ROC curve is 0.804, which is significantly better than the latest method based on the original gray matter / white matter characteristics, and is better than the existing one. Other methods based on Canonical features.
【学位授予单位】：北京理工大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【共引文献】