Recognition Of Pulmonary Nodules On CT Images Based On Stacking Algorithm

In the diagnosis of global cancer,the incidence rate and mortality rate of lung cancer is the top priority.The detection of lung cancer is mainly through the computer tomography(CT)imaging of the lung region to determine whether there is a pulmonary nodule,and then to determine the benign and malignant.However,the problems that massive CT data,the traditional way of manual reading and the lack of experience lead to the doctor’s visual fatigue,resulting in misdiagnosis and missed diagnosis.Therefore,the emergence of computer-aided diagnosis technology(CAD)not only helps radiologists to screen pulmonary nodules effectively,but also improves the treatment level,saves the economic cost and clinical cost.The traditional computer-aided diagnosis technology mainly focuses on the single section of pulmonary nodule,which loses part of the characteristic information of the spatial distribution of pulmonary nodule and makes the accuracy of image recognition of pulmonary nodule low.To solve the problem of incomplete feature extraction of pulmonary nodules,the thesis proposes a texture feature extraction method based on the volume local direction ternary pattern.First of all,the lung nodule slices are arranged in order,and the VLBP model is used to extract the local pattern to form the local pattern.Secondly,based on the normal function,the change of pulmonary nodule image information is self-regulated to form an adaptive threshold,and the local pattern is calculated by ternary.Finally,based on the idea of 3D gray level cooccurrence matrix,the local three pattern formed by threshold adjustment is calculated and normalized to form a 13×3 lung nodule texture feature vector.This method can not only extract the features of pulmonary nodules effectively,but also reduce the feature dimension.In the process of image aided diagnosis of pulmonary nodules,a single classifier can not effectively and accurately classify the unbalanced data and the characteristic vectors of pulmonary nodules.Therefore,the thesis proposes an image recognition method of pulmonary nodules based on stacking evaluation model.In this method,the basic classifier support vector machine,random forest,limit learning machine and k-nearest neighbor classifier are combined to build the stacking evaluation model,and the input data feature analysis,model training and testing,model parameter setting and optimization and model effect evaluation are carried out for the built evaluation model.In the thesis,we use LIDC-IDRI(lung image database)collected by the National Cancer Institute of the United States to carry out the parameter optimization experiment of the volume local direction ternary pattern,the correlation selection experiment of basic classifier and the effect evaluation experiment of lung nodule image aided diagnosis model.The experimental results show that the accuracy,sensitivity,specificity,Matthews correlation coefficient and F1 score of pulmonary nodule recognition based on stacking algorithm are 93.1%,91.8%,90.5%,0.836 and 0.919 respectively.This method has high classification performance and can provide some data reference value for radiologists.