Lung Nodule Recognition Methods Based On Pulmonary High-Resolution CT Images

Lung cancer is now the leading cause of cancer death in the world. Despite advances in treatment and diagnosis, the five-year overall survival rate is only15%. The reason is that lung cancer is usually asymptomatic in early stage and at the time of diagnosis, most patients already present with advanced disease. In order to improve prognosis for lung cancer patients, early diagnosis and treatment would be effective ways. Solitary pulmonary nodule (SPN) is a common manifestation of early lung cancer. Recently, High Resolution Computed Tomography (HRCT) scan has become the primary means for lung cancer screening. However, a large number of image data generated by HRCT will increase the workload for radiologists, which might lead to the increasing risk of missed diagnosis and misdiagnosis. Computer-aided detection (CAD) schemes are effective methods of assisting radiologists in the early detection of cancer in thoracic CT scans, which could help doctors to improve the diagnostic accuracy and reduce the time of film-reading. One of the most important functions for lung cancer CAD is nodule detection. By making an extensive survey on the research status at home and aboard in lung cancer CAD schemes, the purpose of this dissertation is to identify lung nodules in HRCT scans. The contributions of this dissertation are as follows:Firstly, a lung nodule detection method based on analysis of multi-angle enhancement images is proposed. In order to take full advantage of image projection information in three directions and solve the problem that it’s hard to distinguish the lung nodules and pulmonary vascular cross-section of issues in two-dimensional(2-D) image space, nodule enhancement filters are applied on axial plane, coronal plane and sagittal plane to get the corresponding enhanced images, respectively. Then two-dimensional features are extracted based on the enhanced ROIs to eliminate the false positive regions from candidates. Three Support Vector Machines (SVMs) classifiers are trained for three projection planes, respectively. Finally a decision rule is involved to integrate three classification results pixel by pixel. Experiments were performed at32cases with33SPNs. The sensitivity of the recognition method was 92.95%and false positive number was1.04per slice. Experimental results show that this method could locate lung nodules quickly and is superior to the conventional2-D identification methods.Secondly, for lung nodule detection is more precise in three-dimensional (3-D) space, a3-D pulmonary nodule detection approach based on K-L transform and Cost-Sensitive SVM(CS-SVM) is presented. By comprehensively considering the fact that lung nodule identification depends on the accurate segmentation of3-D lung nodules and it’s usually time-consuming for3-D segmentation methods, threshold method and mathematical morphology method are applied on lung fields to obtain the segmentation regions.3-D lung nodule candidates are segmented by using3-D dot-enhancement filter and morphology method. In the false positive reduction stage, intensity features, geometric features and Gabor filter features are extracted based on VOIs and K-L transform are performed for feature selection. In order to solve the problem that the distribution of two samples is unbalanced, CS-SVM is introduced to classify the VOI into nodules and non-nodules. Experimental results show that this approach could identify3-D lung nodules accurately, with high detection sensitivity and low false positive rate.Thirdly, aiming at solving the problem that a large number of false positive regions will be produced by conventional enhancement filter while detecting lung nodules, a3-D multiscale lung nodule enhancement filter based on enhancement density index is developed. A multiscale adaptive bilateral filter is applied to reduce the noisy and smooth CT image sequences. Then, the pre-enhancement coefficients are obtained by computing the Hessian matrix and corresponding eigenvalues pixel by pixel. After analyzing the distribution of pre-enhancement coefficients,3D enhancement density index is constructed. Finally, a discriminant rule based on threshold method is adopted to mark the nodule candidate regions. Experimental results show that this enhancement filter could locate lung nodule region precisely, with high detection sensitivity and few false positive region, which could provide a better initial result for the follow-up processing.Finally, according to a large amount of disturbance caused by pulmonary vascular in the process of nodule detection, a false positive reduction method based on blood vessels elimination is provided. By analyzing and modeling the structure of pulmonary artery regions, a3-D vessel enhancement filter is designed. By using of multiscale approach, blood vesseds with different diameters could be segmented automatically. Finally the procedure of false positive elimination will be applied by matching the enhanced vessel regions and lung nodule candidates. The method was performed on real CT image. Experimental results illustrate the efficiency of the proposed algorithm.