Lung Nodule Detection And Segmentation Methods Based On CT Images

Lung cancer is the highest in the world of male cancer mortality and also high female mortality of cancer. Pulmonary nodule is an early manifestation of lung cancer. CT scanning is the most effective tool for early detection of lung nodules. Computer-aided detection (CAD) system can help radiologists to reduce the time of reading the images and to improve the accuracy of detection. CAD of lung cancer has two main functions: lung nodule detection and lung nodule segmentation. Lung nodule detection comprises candidate nodules extraction and false positive reduction. In this paper, first of all, summarized them in detail, then analyse the shortcoming of existing methods, lastly, proposed the solutions of these problems and achieved good results.First, summarize the lung nodule detection and segmentation methods. Summarize the existing methods according to different types and analysis the existing problem to be solved.Second, aiming at solving the problem that the dot-filter can not extract ground-glass-opacity candidate nodules, a method based on an adaptive nonlinear filter and dot filter was used to extract them. Experiments were performed at 23 scans of CT images (containing 32 GGO nodule).The sensitivity of the detection method was 84.4%, and false positive number was 5.5 per slice. The time cost was 1.2 minutes per scan. This new method is superior to existing methods on the required time, sensitivity and specificity.Third, aiming at solving the problem that the vascular adhesion nodule, the cross point of two or three vessels and endpoint of one vessel can not be distinguished by dot filter and makes high rate of false positive, this paper proposed two new features and a new detection framework based on combination optimization method (improved GA or SS) and SVM. A feature subset selection model based on wrapper model was established, and used improved GA or SS and SVM method to get the best resolution. Using the best feature subset to establish a classifier based on support vector machines to improve the performance by reducing false positive and retaining high true positive. The classifier trained with the optimal feature subset resulted in 80.9% sensitivity and 1.5 false in per scan from the lung nodule data (136 true nodules and 6253 false ones in 64 cases). Experimental results show that the framework and the algorithm are superior to Philips, Siemens and GE's products. Fourth, aiming at solving the problem of bandwidth parameter must be chosen from the long range when Mean-Shift was used in lung nodule segmentation, a new bandwith selectional range reduction method was proposed .Comparing it to the selection method by bandwidth on the basis of statistical analysis, it has the advantages of low complexity at time cost and getting correct bandwidth adaptive to actuality. According to the theorem of bandwidth selection and the regional growing method, the initial parameters of bandwidth were determined, and the most stable scale criterion of multi-scale filtering clustering theory was used to determine the optimal parameters of bandwidth. The proposed method was evaluated and tested for the clinical chest CT images including the nodules of different types, such as the ground glass opacity, the nodules lung walls and vessels adhesion, and solid nodules (anisotropic and isotropic). The test database is made up of 18 scans CT images (36 nodules and 95 slices nodule images). The results reveal that the proposed method is successful in segmentating all types lung nodules.Fifth, aiming at solving the segmentation problem caused by the connection of lung nodule and vessel, a new adaptive bandwidth chosen method based on EM is proposed and apply it into nodule segmentation. Compare it to the method of bandwidth chosen based on statistical analysis rule or optimized rule, it has some advantages such as time low complexity and correct bandwidth accord with real problem. Imposing the flow feature orientation vectors of vessel submitting to normal distribution and the flow feature orientation vectors of nodule submitting to uniform distribution, modeling the nodule connected vessel, and estimating model parameter by EM, extracting bandwidth values in Mean-Shift based on the weight of uniform distribution and bandwidth selection theorem. The proposed method was tested on the 16 scans clinical chest CT images, and all the results are correct. This method provides a powerful tool for 3D segmentating lung nodules connected vessel.