Reasearch On Classification Of Benign And Malignant Pulmonary Nodules Based On CT Images

Lung cancer has become one of the most fatal malignant cancers in the world,which is a serious threat to the public health.The key point on enhancing the cure rate and reducing the death rate is to diagnose lung cancer in the early stage.Computed Tomography(CT)is believed to be the best imaging method in clinical for lung cancer diagnosis.However,manually analyzing CT images is always time-consuming and human-fatigued,and this will result in the high misdiagnosis and missed diagnosis rate.Nodule,as the most common feature of lung cancer,can be diagnosed by using computer-aided diagnosis system(CAD),which helps radiologists to reduce the workload and improve the precisely diagnosis rate.This thesis focuses on key techniques reasearch of computer-aided diagnosis based on CT images and they cover the construction of lung nodule database,lung nodule feature learning,and classification of benign and malignant lung nodules.The main contributions of this thesis are listed below:(1)Construction of gold standard database for lung nodules.The construction of lung nodule gold standard database is the key to train and test the computer-aided diagnosis system for lung nodules.Lung nodule dataset in this thesis is extracted from the world’s largest chest image database LIDC-IDRI.Aiming at solving that each nodule is annotated by four radiologists independently,this thesis obtains accurate segmentation of a lung nodule by fusing the four boundaries of this nodule with STAPLE algorithm.In addition,benign and malignant pulmonary nodules can be determined according to the rating scales of each lung nodule.(2)Considering the visual similarity and semantic relevance of lung nodules,a classification algorithm based on similarity metric is proposed.A medical image retrieval model is constructed based on this similarity metric method,which can be used to search for similar nodules in the reference dataset.The query nodules are then diagnosed benign or malignant on the basis of the retrieval results.The retrieval nodules from this model are not only semantic relevance,but visually similar to the query nodules,which can provide a reference for the doctor’s diagnosis.The experimental results demonstrate the feasibility and effectiveness of the proposed algorithm,especially the retrieval accuracy is significantly superior to the existing measurement algorithms.(3)Aiming at the clustering of unlabeled pulmonary nodules,a new spectral cltustering algorithm is explored for clustering of benign and malignant lung nodules.A new laplacian matrix is constructed by the local kernel regression model,which can be used to cluster the in-sample data.Aiming at the clustering of out-of-sample data,a linear regression regularization term or a global kernel regression regularization term is introduced to the new laplacian matrix.These regularization terms can be used to cluster each out-of-sample data.The experimental results show that the proposed algorithm outperforms the existing clustering algorithms.(4)Lung nodule feature learning problem.According to the doctor’s experience of understanding CT image features and diagnosis of benign and malignant pulmonary nodules,two features of lesion density level and heterogeneity are proposed to represent nodules,and a two-step similarity metric method is used to classify lung nodules.A new lung nodule database is constructed to verify the robustness of the proposed method.Experimental results demonstrate the feasibility and robustness of the proposed method.Aiming at solving the classification problem of five malignant ratings of lung nodules marked by radiologists,a new k nearest neighbor(kNN)scheme for identifying lung nodule malignant ratings is proposed,which further demonstrates the feasibility of the pulmonary nodule feature learning.(5)Studying on diagnosis of benign and malignant pulmonary nodules based on texture features and similarity measurement.In this research,we study the feature learning and similarity measurement in medical image retrieval From the perspective of feature learning,three different types of texture features of pulmonary nodule images are studied;from the perspective of similarity measurement,a two-step similarity measure of medical image retrieval scheme is proposed to assist in the diagnosis of benign and malignant pulmonary nodules.The experimental results show that Haralick texture features are more effective in distinguishing benign and malignant pulmonary nodules.Our proposed medical image retrieval scheme is feasible for the classification of benign and malignant pulmonary nodules.