Research On Pulmonary Nodule Detection Based On Lung CT Images

Lung cancer is one of the highest incidence and mortality in malignant tumors worldwide.Early detection,early diagnosis and early treatment is of great practical significance.The computer aided detection(CAD)systems for potentially cancerous pulmonary nodules based on CT images have become considerably crucial to improve the patients’relative survival rate.CAD systems could not only enhance the accuracy and sensitivity of nodule detection,but also significantly reduce physician time needed for interpretation and negligence caused by radiologists’unilateral reading CT slices.It is a“second option”to assist radiologists making final decisions.Given the complexity and heterogeneous of lung cancer,based on the CAD systems for pulmonary nodule detection in CT images,the subject“Research on pulmonary nodule detection for lung CT images”is proposed.Techniques in those key stages of feature extraction,feature dimensionality reduction and classification in a CAD system are deeply analyzed,improved and investigated experimentally,so as to further enhance the performance of the CAD system.The main research contents and innovations are as follows:(1)Research on dimensionality reduction technique of pulmonary nodule features based on manifold learning method.Since the complexity of lung CT images determines the high dimensional and non-linear characteristics of the feature space,the supervised locally linear embedding(SLLE)algorithm based on manifold learning is adopted to reduce the dimensionality.However,the Euclidean distance in SLLE could not well represent the similarity between data points in high dimensional space,thus a Spearman correlation coefficient based SLLE algorithm,named SC~2SLLE,is presented.By introducing the Spearman correlation coefficient,a new distance metric is constructed,which is suitable for measuring the similarity between data points in high dimensional space,improving the Euclidean distance in SLLE.The experiment is conducted by using the lung CT images collected from the international standard LIDC/IDRI database.After image preprocessing,pulmonary parenchyma segmentation and pulmonary nodule segmentation,all nodule candidates are extracted,and then the gray features,texture features and morphological features are extracted.The proposed SC~2SLLE algorithm is applied to reduce the dimensionality of the features.Finally,the support vector machine(SVM)combined with the low dimensional features are used for classification.Experimental results show that the proposed SC~2SLLE algorithm is superior to the SLLE and the unsupervised LLE algorithm.Prominent dimensionality reduction results are achieved by the SC~2SLLE algorithm.(2)Research on dictionary learning and spare representation based feature extraction method of pulmonary nodules.Given the problem that these current segmentation methods on lung CT images are difficult to obtain precise results,a new feature extraction approach based on dictionary learning and spare representation framework is proposed.Namely,a discriminative class-specific dictionary is learnt for each class of the training samples and a background dictionary containing sharing information are learnt for all training samples,based on the classification-oriented dictionary learning model.Then a big dictionary is constructed with these class-specific dictionaries and the background dictionary.The orthogonal matching pursuit(OMP)algorithm is used for spare representation upon the big dictionary.In terms of the correspondences between the dictionary and the sparse representation,as well as the contribution level of each atom to the spare representation of each sample,the distinctive features are extracted from the coefficient matrix.The national standard LIDC/IDRI database is adopted to conduct experiment for the detection of solitary and juxta-pleural pulmonary nodules.First the train samples are used to train two class-specific dictionaries and one background dictionary,then feature dataset is extracted upon the proposed dictionary learning based feature extraction approach.After dimensionality reduction by using the SC~2SLLE algorithm,the SVM classifier is adopted for classification.Experimental results demonstrate that the proposed feature extraction method is effective and feasible,whose AUC value(the area under the ROC curve,an indicator of measuring the diagnosis experiment comprehensively)reaches 0.9041,which sufficiently manifests that the features extracted from the dictionary learning based feature extraction method is more representative,since it reflects the intrinsic characteristics of the data samples.(3)Research on improved relevant vector machine(RVM)method based pulmonary nodule detection.Because the setup of the kernel parameters in RVM is empirically determined or selected by uniform distribution selection method and so on,and single kernel based RVM has limitations in dealing with complex dataset and kernel function selection,thus in order to alleviate those issues,the particle swarm optimization(PSO)algorithm and the second-order cone programming(SOCP)algorithm based multi-kernel learning RVM method,named PSO-SOCP-MKLRVM,is developed.First the PSO algorithm is adopted to optimize the kernel parameter in a single kernel RVM model.Based on this,the multi-kernel RVM model is constructed,and then it is derived and proved that the optimization problem of multi-kernel alignment RVM model is a convex problem,hence the SOCP is employed to quickly calculate the combination coefficients between different kernel functions.Thus,the PSO-SOCP-MKLRVM model is constructed.The experiment of pulmonary nodule detection on the LIDC/IDRI database is conducted.Namely,the dictionary learning model based feature extraction method is used to extract features,the SC~2SLLE algorithm is adopted to reduce dimension,and the PSO-SOCP-MKLRVM method is employed for classification.Experimental results demonstrate that the proposed multi-kernel classification method outperforms the current multi-kernel RVM and single-kernel RVM methods.The AUC value is up to 0.9117,apparently the performance of the proposed classification method is remarkable in pulmonary nodule detection,which provides scientific evidence for precise detection of pulmonary nodules.