Research On Key Technologies Of Computer Aided Diagnosis For Malignant Pulmonary Nodules

Currently,lung cancer is one of the malignant tumors with the highest morbidity and mortality worldwide,which seriously threatens human health and life safety.Extensive clinical practices demonstrate that focus on early detection,early diagnosis and early treatment is one of the most effective means of decreasing mortality rates and improving five-year survival rates.Lung cancer at the early stage is mainly manifested as pulmonary nodules,therefore the early diagnosis of pulmonary nodules is of great importance for diagnosis and treatment of early lung cancer.Recently,with the rapid development of computer technology and medical imaging technology,computer-aided diagnosis for malignant pulmonary nodules has gradually become one of hot researches,and has the high value in clinical application.CAD not only reduces the workload of the clinicians,but also improves the accuracy,and increases the risk of misdiagnosis or diagnostic errors for the diagnosis of benign and malignant pulmonary nodules.In this paper,through the in-depth research and analysis on the key technologies of the existing CAD for malignant pulmonary nodules,a series of improved algorithms are proposed in this paper.The main contributions of this paper can be summarized as follows:(1)To address GGO pulmonary nodules segmentation problem of low accuracy caused by some factors,such as fuzzy boundaries,irregular shapes,inhomogeneous intensities and low contrast,through the in-depth research on the traditional random walker algorithm,a modified random walker algorithm is proposed in this paper.To enhance the pulmonary nodules,a novel multiscale dot enhancement filter based on the eigenvalues of Hessian matrix and shape index is proposed in this paper.To automatically acquire seed points located in pulmonary nodule regions,the enhanced pulmonary nodules are thresholded.The shape index and Gabor texture feature are combined to select seed points located in background regions.To address the traditional random walker segmentation problem of low accuracy caused by depending on intensity information,intensity feature,texture feature and spatial location information of pulmonary nodules are combined to construct a new weighted function,which more effectively measures the similarity relationships between pairwise 8-neighboring nodes.The label restriction energy term is added into the energy function of traditional random walker model.The algorithm effectively use the initial label information of seed points,thus improves the accuracy and efficiency of GGO pulmonary nodule segmentation.(2)Through the in-depth research on the traditional graph construction methods,some problems are pointed out,such as the difficulty of parameter selection and the sensitivity of noise.To address these problems,a segmentation algorithm based on sparse representation and random walk is proposed for segmentation of GGO pulmonary nodules in this paper.To enhance the pulmonary nodules,a new multiscale dot enhancement filter based on shape index and curvature is proposed.Geodesic distance method is used to select the initial seed points located in nodule regions,and a local searching strategy is presented to select the other seed points located in nodule regions and seeds points located in background regions.The 8-neighboring weighted function and the weighted function based on sparse representation coefficients and k-NN method are constructed to more effectively measure the similarity relationships between pairwise neighboring nodes.This new weighted graph avoids the pseudo-similarity caused by noise interference.Finally,a new energy function is defined,including 8-neighbor data penalty term,sparse representation and k-NN data penalty term,and label restriction energy term.The experimental results demonstrate that the algorithm further improves the accuracy and efficiency of GGO pulmonary nodule segmentation.(3)Through the in-depth research on the problems of the feature extraction and classification methods,the limitations of methods are pointed out.A classification algorithm based on radiomics and random forests algorithm is proposed for classifying benign and malignant pulmonary nodules in this paper.To address the problem of incomplete features in existing feature extraction methods,a radiomics method is used to mining high-demensional pulmonary nodule features,including intensity features,geometric features and texture features.Thus,these features comprehensively describe the characteristics of pulmonary nodules.Gray level co-occurrence matrix method,local binary patterns method and Gabor filter-based method are combined to extract texture features.Thus,these texture features describe the differences between benign and malignant pulmonary nodules.To address pulmonary nodules classification problem of low accuracy caused by the simple classifiers,a random forest algorithm based KL-divergence is proposed.The final output classification results are decided by the weighted votes of each decision.The experimental results demonstrate that the algorithm improves the accuracy and stability of benign and malignant pulmonary nodule classification.(4)To address pulmonary nodule classification problem of low accuracy caused by a small number of labeled pulmonary nodule samples,through the in-depth research on the semisupervised learning and label propagation algorithms,some problems are pointed out,such as the sensitivity of noise and the lack of local information among other sample data.To address these problems,an iterative label propagation is proposed for classification of benign and malignant pulmonary nodules.A k-NN graph based on k-NN method is constructed to better capture the connected relationships between the neighboring samples.The geodesic distance instead of Euclidean distance is used to define a weighted function,which more effectively measures the similarity relationships between pairwise neighboring samples.Finally,a cost function is defined by considering the effect of the labels of other k-NN sample data.The labels of labeled sample data can be correctly propagated to the unlabeled sample data in an iterative way.Thus,the algorithm improves the accuracy and stability for classification of benign and malignant pulmonary nodules in the case of small pulmonary nodule samples.The experimental results demonstrate that the improved algorithms achieve the high accuracy,sensitivity and specificity for classifying benign and malignant pulmonary nodules.Therefore,the algorithms plays a supporting role for physician’s discriminate diagnosis of pulmonary nodules.However,it is necessary to further optimize and modify the segmentation and classification algorithms in order to meet the clinical applications.