Research And Realization Of Detection And Classification Of Pulmonary Nodules Based On CT Image Features

Lung cancer is one of the most concerned cancer diseases in the scientific community and its high morbidity and mortality rates make researchers maintaining researches on lung cancer.Early detection and diagnosis of lung cancer is an effective way to alleviate moridity and reduce mortality.The development of computer tomography(CT)technology has provided researchers with high quality pleural CT images and more image details have been highlighted.In addition,the accumulation of a large amount of clinical medical image data gives researchers an opportunity to further explore the pathologic features of lung cancer.The existing computer-aided detection and diagnostic(CAD:CADe/CADx)system can effectively relieve the radiologist's heavy manual inspection and diagnostic stress.Therefore,CAD system has become an indispensable clinical technical program.However,some weaknesses and limitations remain in the current CAD systems.To improve the performance of current CAD systems and provide valuable research methods for this field,this dissertation based on high-resolution lung CT images,for lung nodule detection,diagnosis and pulmonary pathologic assessment,has accomplished the following research work:(1)This dissertation collects the largest and universally accepted chest CT image database LIDC-IDRI which contains pulmonary nodules and its detailed notes.The database covers 1018 sets of CT images of high-risk lung cancer patients and detailed pulmonary nodular annotation information of each case can be refered in the XML file respectively.For this XML file,this dissertation has designed a framework that contains all the pulmonary nodule attributes and annotation information,laying the foundation for further testing and diagnosis.In addition,this study,in collaboration with a medical school in Guangzhou,analyzed the CT images of 115 patients with lung cancer and performed a tracking test,which provided clinical guarantee for this study.(2)In the study of lung segmentation and pulmonary nodule selection,in order to unbiasedly divide the parenchymal region and divide the suspected pulmonary nodule region in the lung,this dissertation presents a segmentation algorithm based on three-dimensional regional growth and a candidate lung noudles selection algorithm based on space fuzzy C-means(SFCM)with Gaussian kernel function.Moreover,an automated pulmonary contour repair algorithm based on contour differential analysis in sliding windows is proposed.Experiments demonstrated that this method has a good generalization in dealing with a large amount of lung CT images with high abundant noudle information.(3)In the study of pulmonary nodule identification and diagnosis,this dissertation mainly focuses on reducing the false positive pulmonary nodules and improving the feature extraction method.At first,a frangi filter based on Hession matrix is used to eliminate the tubular false positive nodule region.Secondly,this study presents a multi-group and multi-scale pulmonary nodule recognition method based on depth learning.Finally,based on the segmented pulmonary nodule region,this dissertation combines a variety of statistical features,texture features and shape features and proposed pixel gray value of the statistical data(PVSSM)and three-dimensional rotation invariance of the LBP feature novelly.By combining a variety of pattern recognition algorithms,the benign and malignant attributes of pulmonary nodules can be judged.Experiments show that the method can effectively reduce the false positive rate of nodules and prove the feasibility of using PVSSM feature to make assessment on benign and malignant pulmonary nodules.(4)In the study of pulmonary nodal diagnosis and pathology assessment of pulmonary parenchyma,this dissertation presents an analytical method for the potential Tirichlet distribution(LDA)subject model based on image words.First of all,this dissertation designs the method of constructing two-dimensional and three-dimensional image words based on patch respectively.Secondly,through the multi-dimensional feature extraction of two-dimensional and three-dimensional images,each image word is expressed as one-dimensional eigenvector.Then,the LDA theme model based on image corpus is constructed,and the number of image vocabulary and subject is determined based on empirical and experimental analysis.Finally,the study through the Gibbs sampling method trains the LDA theme model and performs pathological analysis of pulmonary nodules and lung tissue with statistical analysis methods.Experiments indicate that this method can achieve multi-attribute analysis of pulmonary nodules and lung tissue with significant analysis results.In this dissertation,a series of algorithms on pulmonary nodule detection and pulmonary nodule and lung tissue diagnosis system can not only promote the clinical application of lung cancer CAD system,but also provide a novel solution for lung cancer detection and diagnosis.