基于脑部磁共振图像的特征提取与分类

CASIA OpenIR > 毕业生 > 硕士学位论文

	基于脑部磁共振图像的特征提取与分类
	李涛
	2016-11-23
学位类型	工学硕士
英文摘要	脑疾病严重威胁着人类的健康，脑疾病的早期诊断对疾病治疗以及相关药物的研究，都具有十分重要的意义。利用磁共振结构图像，提取脑结构相关特征，建立更为准确的分类模型，辅助医疗诊断成为当前研究的热点。传统的磁共振结构图像分类算法，多基于图像的底层特征，如灰白质密度、脑区容积、皮层厚度和形状特征等。然而，这些底层特征具有一定的局限性：1）底层特征的精确提取依赖于样本脑图像间的配准结果，由于缺乏金标准，配准结果难以评估，容易造成误配准。2）一些结构差异容易在配准过程中被消除。为了解决上述问题，本文基于词袋模型、局部特征融合和自编码神经网络，提出了三种磁共振结构图像分类算法，包括基于多级分区词袋模型的脑疾病识别方法、两阶段局部特征融合的脑疾病识别算法和基于三维自编码神经网络的脑疾病识别方法。本文的创新性研究工作主要有：（1）词袋模型是一种有效的图像表示方法，针对词袋模型缺乏对特征位置信息的有效利用，本文提出了基于多级分区词袋模型的脑疾病识别方法。本方法基于标准脑模板，采用多级脑分区划分，构建各等级分区词袋，进而建立各等级词袋直方图对图像进行表示，利用分类器集成方法对测试样本分类。本方法旨在利用不同尺度的特征分布构造词袋，进行磁共振图像分类和个体属性判定。在标准数据集OASIS上的实验结果表明：相比于不分区，多级脑分区能够综合各脑分区的特征分布信息，建立更为有效的分类模型，在正确率、精确度、敏感性和特异性四个指标都体现了相对优势。（2）针对底层特征提取需要样本脑图像之间精确配准的问题，本文提出一种两阶段局部特征融合的磁共振结构图像分类算法。在结构匹配阶段，首先在尺度空间提取关键点，使用SIFT描述子将不同样本同一组织结构（关键点）互相匹配。在结构描述阶段，使用HOG特征将用以细化关键点附近的形态特征，区分不同人群。本方法旨在将两种局部特征融合，在不精确配准的情况下准确分类不同人群。在标准数据集OASIS和PPMI上，测试该方法的性能。（3）针对底层特征表达能力有限，临床指标在现有分类方法中并未有效利用，本文提出了基于三维自编码神经网络和多任务学习的脑疾病识别方法，克服特征提取中的局限性并综合利用阿尔茨海默病人的临床测试指标。本方法首先应用自编码神经网络对深度神经元网络做非监督的分层预训练，将得到的参数作为神经网络模型的初始化参数；然后使用有标签的数据对网络微调完成监督训练，得到整个神经网络的模型参数。本方法旨在利用深度学习神经元网络的自学习方法，提升特征的表示能力，使学到的特征更加抽象和反应原始数据的本质。此外，本方法希望利用多任务学习综合利用临床测试数据，实现更好的分类。在标准数据集ADNI上测试该方法的性能，并表明神经网络模型具有较强的特征学习能力。; As brain disease threaten human's health very much, the early diagnose of brain disease can offer effective power to detect the earliest sign of illness and resolve drug effects in clinical trials. Growing interest has been focused on the extraction of structural features and the accurate classification of brain disease using magnetic resonance images. The traditional classification methods are based on low-level features such as voxel-wise density, cortical thickness, volume, and deformation information. However, there are several constraints in such features: 1) To extract these features precisely, the brains require nonlinear alignments to a template. For the lack of ground truth, it's hard to evaluate the performance of inter-subject registration. There is the risk of misalignment or over-alignment. 2) The differences caused by disease may be removed in the registration process. To overcome the above limitations, we propose three classification methods based on bag-of-words (BOW), local feature fusion, and autoencoder neural network: classification of brain disease from magnetic resonance images based on multi-level brain partitions, classification of brain disease in magnetic resonance images using two-stage local feature fusion, classification of brain disease in magnetic resonance images via autoencoder neural network. (1) BOW is a effective model for image representation. As the location information of features are ignored in BOW, we provide a classification method based on multi-level brain partitions. The template brain is divided into multi-level partitions, then bags are constructed for brain partitions. BOW histogram, a vector of word's frequency, is calculated to represent MR image and used to classify new subject. This method is proposed to utilize more spatial information of different brain regions. The results evaluated on public dataset OASIS show that: compared with BOW model based on the whole brain, BOW model based on multi-level partitions obtain better performance in terms of accuracy, precision, sensitivity and specificity. (2) To extract these features precisely, accurate alignments are required among subjects. To solve this constrain, we propose a classification method using two-stage local feature fusion. In correspondence stage, we extract keypoints and SIFT descriptor from scale-space in MR images. SIFT descriptors are used to correspond same anatomical structure among different subjects. In representation stage, HOG descriptors will be calculated to demonstrate the local character around the representative keypoints from the same scale. This method aims to classify new subjects without precise registration via the fusion of two kind local features. We evaluate the performance of our method on public dataset OASIS and PPMI. (3) As the limitation of low-level features in representation and the ineffective utilization of clinical indices, we propose a classification method based on 3-D convolution autoencoder neural network and multi-task learning. 3D network is built upon a convolution autoencoder, which is pre-trained to capture the structural feature of MR images. Then a full connected layer is then fine-tuned for classification task. This method aims to utilize the powerful effect of deep learning framework to improve the representation ability of features. The features we obtain from deep learning framework may demonstrate the nature of original data. Besides, we hope to utilize multi-task learning to integrate different kinds of clinical indices. We evaluate our method on public dataset ADNI, and the result validate the ability of network in feature learning.
关键词	脑疾病局部特征 Sift特征词袋模型神经网络
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/12806
专题	毕业生_硕士学位论文
作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	李涛. 基于脑部磁共振图像的特征提取与分类[D]. 北京. 中国科学院研究生院,2016.

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
基于脑部磁共振图像的特征提取与分类.pd（2180KB）	学位论文		限制开放	CC BY-NC-SA