基于特征融合的多模MRI脑肿瘤分割

发布时间：2018-01-03 15:11

本文关键词：基于特征融合的多模MRI脑肿瘤分割　出处：《武汉理工大学》2015年硕士论文　论文类型：学位论文

【摘要】：在医学影像中,核磁共振成像(Magnetic Resonance Imaging,MRI)是一种重要的成像技术,此成像技术有着高质量的图像显示效果,已被广泛应用于医学中对人体各组织器官病变的诊断中,特别是对脑部病变组织的检测。MRI脑肿瘤分割在实际的临床诊断上提供了很大的帮助,如何能够更快速准确的分割出脑肿瘤是MRI脑肿瘤研究的难点。又因脑肿瘤核磁共振图像的可变性和复杂性,以及脑肿瘤的大小、形状都各不相同,对脑肿瘤特征的提取就显得尤为重要。近年来,关于此方面的研究也有很多,虽然有很不错的效果,但是在此研究领域还有广泛的提升空间。本文主要研究MRI脑肿瘤分割:将Gabor小波提取的脑肿瘤特征与卷积神经网络提取的脑肿瘤特征用特征融合方法形成新的特征,然后在此改进算法的基础上引入了一种新的降维方法——核熵成分分析算法对融合后的特征进行降维,并取得了不错的效果,具体工作内容如下:(1)分析了Gabor小波算法的基本原理,使用5个尺度和8个方向的40个Gabor滤波器进行卷积,提取特征,并将卷积后的结果作为特征向量,然后验证了基于SVM(Support Vector Machine)的脑肿瘤分割,并对分割结果做后处理,将其在GBM数据集上进行实验,并对结果进行分析。(2)分析了卷积神经网络的基本原理,研究了卷积神经网络的网络结构和训练过程。在卷积层作卷积运算,增强了原始信号强度,降低了噪声;在降采样层对前层图像作抽样处理,此方法在保证不降低有用信息的基础上减少了需要处理的数据量。在卷积神经网络中还使用了参数减少和权值共享等方式,提高了运算速度。将其应用到GBM数据集上进行实验,并分析其优势和劣势。(3)为了提高分割的精度,提出了将人工选取的特征与机器学习的特征相结合的方法。展开了特征融合的相关工作,即将Gabor小波和卷积神经网络提取的特征根据特征融合中的串行组合方式串连成一个列向量作为新的特征向量,同时基于核熵成分分析对融合后的新特征作降维处理,然后对降维前与降维后的特征分别用SVM进行分类,根据实验结果分析其优劣。
[Abstract]:Magnetic Resonance imaging (MRI) is an important imaging technique in medical imaging. This imaging technique has a high quality image display effect, and has been widely used in the diagnosis of human tissues and organs in medicine. In particular, the detection of brain lesions. MRI brain tumor segmentation in the actual clinical diagnosis provides a great help. How to segment brain tumors more quickly and accurately is a difficult point in the study of brain tumors in MRI. Because of the variability and complexity of MRI images and the size and shape of brain tumors are different. The extraction of brain tumor features is particularly important. In recent years, there are a lot of research on this aspect, although there are very good results. However, there is still wide scope for improvement in this field. This paper focuses on the segmentation of MRI brain tumors:. The features of brain tumors extracted by Gabor wavelet and those extracted by convolution neural network are fused to form new features. Based on the improved algorithm, a new dimensionality reduction method, kernel entropy component analysis algorithm, is introduced to reduce the dimension of the fused features, and good results are obtained. The main work is as follows: (1) the basic principle of Gabor wavelet algorithm is analyzed, and 40 Gabor filters with 5 scales and 8 directions are used for convolution and feature extraction. The convolution result is taken as the feature vector, and then the segmentation of brain tumor based on SVM(Support Vector Machine is verified, and the segmentation result is processed. The basic principle of the convolutional neural network is analyzed by the experiment on the GBM data set and the analysis of the results. The network structure and training process of the convolution neural network are studied. The convolution operation in the convolution layer enhances the intensity of the original signal and reduces the noise. This method can reduce the amount of data needed to be processed on the basis of not reducing the useful information, and using parameter reduction and weight sharing in the convolution neural network. In order to improve the segmentation accuracy, the algorithm is applied to the GBM data set for experiment, and its advantages and disadvantages are analyzed. In this paper, a method of combining artificial selected features with machine learning features is proposed, and the related work of feature fusion is carried out. The feature extracted by Gabor wavelet and convolution neural network is serially connected to a column vector as a new feature vector according to the serial combination in feature fusion. At the same time, the new features are reduced based on kernel entropy component analysis, and then the features before and after dimensionality reduction are classified by SVM, and the advantages and disadvantages are analyzed according to the experimental results.
【学位授予单位】：武汉理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：R739.41;R445.2;TP391.41

【参考文献】