Lung CT Images Recognition By Low Rank Approximation And Support Vector Machine

Lung cancer is one of the most malignant tumors that are most harmful to human beings.The early form of lung cancer is lung nodules.Therefore,the identification of lung nodules is very critical for finding lung cancer and improving the survival rate as early as possible.The detection accuracy of lung nodules in Computer-Aided Detection(CAD)system is closely related to the performance of the feature vector extraction module,data reduction module and classification module.Aiming at the heterogeneity and complexity of lung nodules,this thesis tries to find new feature extraction methods and classifier optimization algorithms so as to improve the classification accuracy of the lung nodule CT imagesThe main contents of this thesis are as follows:(1)The improvement of optimization algorithm used for the support vector machine model.The bat algorithm(BAT)is introduced to optimize the parameters of SVM.In view of the characteristic that the bat algorithm is easy to fall into the local optimum,the cloud model is used to improve the bat algorithm in order for better global-andlocal search ability.and then the new support vector machine algorithm based on the cloud model and bat algorithm(CBAT-SVM)is proposed.The classical test functions experiments show that the CBAT algorithm is superior to the BAT algorithm in accuracy.The tests of UCI datasets indicate that the CBAT-SVM algorithm has better classification performance.(2)The improvement of the feature extraction method for lung nodules.Because of the specificity of lung nodules,the selected feature vectors will directly affect the classification results of classifiers.Due to the obvious heterogeneity,texture and complexity of the lung nodules,this thesis adds the Curvelet transform coefficients into the feature vectors on the basis of the conventional feature extraction method in order to improve the morphological and textural characteristics of the lung nodules.The experimental results of lung tissue slices show that Curvelet transform has better performance in expression of the morphological and textural features of images.(3)Study on dimensionality reduction methods using low rank approximation for lung nodule feature vectors.In order to reduce the pressure caused by excessive eigenvector dimensions of lung nodules during the learning process with support vector machine,this thesis studies the dimensionality reduction effect of three dimensionality reduction methods using low rank approximation on the feature vectors of lung nodules,such as Principal Component Analysis(PCA),Linear Differential Analysis(LDA)and Neighborhood Preserving Embedding(NPE).The experimental analysis shows that the combination of PCA and LDA is more suitable for dimensionality reduction of new lung nodule feature vectors in which the Curvelet transform coefficients are added.(4)The classification and application of lung nodule CT images.By studying the methods and algorithms of feature vector extraction,data dimensionality reduction and classification of lung CT images,a lung nodule auxiliary detection system suitable for CT images is designed,which includes improved feature extraction module,suitable dimensionality reduction module with low rank approximation methods and support vector machine module based on CBAT algorithm.The detection experiments conducted with LIDC/IDRI database show that the recognition accuracy of the designed detection system is satisfactory,and the designed detection system is far superior to the conventional recognition methods,which provides a scientific basis for accurate detection of lung nodules.