基于特征选择的医学图像分类方法及其应用研究

发布时间：2019-06-29 14:18

【摘要】：随着计算机视觉技术及医疗成像设备的不断发展,医学图像包含的信息也极为丰富,所以医学图像的分类识别技术越来越重要,而图像特征的提取和选择作为图像分类识别的关键环节,使得国内外学者越发关注。图像处理技术的完善导致处理图像得到的信息特征的维数也越来越高,因此选取更有效的特征对图像进行分类处理就成为关键。本文主要研究类风湿性关节炎(Rheumatoid Arthritis,RA)医学图像,以往医学图像提取的特征大多数是基于图像的纹理、颜色和形状等,而本文是对RA患者手指的扩展光学层析图像(Diffuse Optical Tomography,DOT)提取光学系数作为特征。通过对RA图像的分类实验,验证了光学系数作为分类特征的可行性。特征选择是医学图像诊断的关键方法,在研究了众多特征选择方法之后,对最大相关最小冗余(Maximal Relevance and Minimal Redundancy,MRMR)特征选择算法进行深入研究,此算法将特征的相关性和冗余性相结合,并利用熵运算对特征排序,选择出最佳的特征组合。继续研究发现MRMR算法在相关性和冗余性有不同权重占比的情况下,特征的排序会有所不同,因此加入了权值进行研究,特征序列采用降序,之后对不同权值下的特征序列按照位置的权重系数再一次排序处理,得到最终的特征序列,依照此序列进行分类实验,心脏病和RA诊断分类实验都证明了改进的MRMR算法的确实有效性。在进行医学图像分类之前,为保证分类的准确率,需要对样本进行筛选,而且病症诊断的速度对患者非常重要,所以寻求小样本学习方法就至关重要。支持向量机(Support Vector Machine,SVM)是基于结构风险最小理论的适用于小样本学习的最佳学习方法,有很好的泛化能力,又能有效的减小训练误差,而在建立分类模型的过程中,会遇到参数和核函数选择的困扰,本文采用基于密度的的聚类分析算法,结合粒子群优化算法(Particle Swarm Optimization,PSO)对分类器进行优化,将繁多的数据集进行精简,进而有效的提高分类器的分辨准确率。仿真实验结果表明该方法行之有效,并且对特征众多需要合理筛选的医学图像处理有参考价值。
[Abstract]:With the continuous development of computer vision technology and medical imaging equipment, the information contained in medical images is also very rich, so the classification and recognition technology of medical images is becoming more and more important. As the key link of image classification and recognition, the extraction and selection of image features make scholars at home and abroad pay more and more attention. With the improvement of image processing technology, the dimension of the information features obtained from the image processing is getting higher and higher, so it is the key to select more effective features to classify and process the image. In this paper, we mainly study the medical images of rheumatoid arthritis (Rheumatoid Arthritis,RA). In the past, most of the features extracted from medical images are based on the texture, color and shape of the images, while this paper is to extract the optical coefficients of the extended optical tomography image (Diffuse Optical Tomography,DOT of the fingers of RA patients as features. The feasibility of optical coefficient as a classification feature is verified by the classification experiment of RA images. Feature selection is the key method of medical image diagnosis. After studying many feature selection methods, the maximum correlation minimum redundant (Maximal Relevance and Minimal Redundancy,MRMR (maximum correlation minimum redundant feature selection algorithm) feature selection algorithm is deeply studied. This algorithm combines the correlation and redundancy of features, and uses entropy operation to sort the features and select the best feature combination. It is found that the feature ranking of MRMR algorithm will be different when there are different weights in correlation and redundancy, so the feature sequence is studied by adding weights, and then the feature sequence under different weights is sorted again according to the weight coefficient of position, and the final feature sequence is obtained, and the classification experiment is carried out according to this sequence. The classification experiments of heart disease and RA diagnosis show the effectiveness of the improved MRMR algorithm. Before medical image classification, in order to ensure the accuracy of classification, it is necessary to screen the samples, and the speed of disease diagnosis is very important for patients, so it is very important to find a small sample learning method. Support vector machine (Support Vector Machine,SVM) is the best learning method based on the minimum theory of structural risk, which has good generalization ability and can effectively reduce the training error. In the process of establishing the classification model, it will encounter the problem of parameter and kernel function selection. In this paper, the density-based clustering analysis algorithm combined with particle swarm optimization algorithm (Particle Swarm Optimization,PSO) is used to optimize the classification. A large number of data sets are simplified to effectively improve the resolution accuracy of the classifiers. The simulation results show that the method is effective and valuable for medical image processing, which needs reasonable selection of features.
【学位授予单位】：长春工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】