Studies On Computer-aided Diagnosis For Pulmonary Nodules From CT Image

Lung cancer,as a malignant tumor with high morbidity and mortality,has serious threat to human life and property safety.Early manifestation of lung cancer is pulmonary nodule.Pulmonary nodule is quasi-circular lesion whose diameter between 3 to 30 mm in the lung.The benign and malignant judgment of pulmonary nodule is a great significance in the early diagnosis of lung cancer.computed tomography(CT)is a common imaging way to check pulmonary nodule,and its high spatial resolution can clearly show the lung tissue.Computer-aided diagnosis(CAD)is a kind of technology,which using the method of pattern recognition and medical image processing,combined with the efficient data analysis capability of computer to assist the doctor in finding and diagnosing the lesion.In this study,CAD for pulmonary nodules based on chest CT image from Liaoning Cancer Hospital & Institute will be studied,the main work are as follows:(1)Pretreatment of chest CT image: the CT images were denoised by median filtering,and the parenchyma was divided by threshold method and morphological method.In this study,a method of repairing pulmonary parenchyma based on convex hull algorithm and curve fitting is proposed to solve the problem that the lung parenchyma can’t fully divided into the lung parenchyma.It is proved that this method can effectively repair the abnormal depression area on the basis of dividing the initial pulmonary parenchyma,so as to ensure that the pulmonary wall adhesions are not missed when the pulmonary parenchyma is divided.(2)Computer-assisted of true and false recognition for pulmonary nodules : basing on the isolated lung parenchyma,we can use threshold method to segment the suspected nodular region.In order to reduce the impact of subsequent classification of lung nodules due to inaccurate classification,for the vascular adhesions,we correct the nodular area to remove a large number of blood vessels by using the blow the ball method.The area of suspected nodules is divided into training set and test set,and the data of the training set is augmented by a random rotation.The convolutional neural network(CNN)is trained by the training set with category.The trained model is used to predict the test set and output the test scores of the true or false of each sample,the results showed the high accuracy rate,the high true positive rate and the low false positive.The experimental results showed that CNN was better for computer-assisted of true or false recognition for pulmonary nodules.(3)Computer-assisted of benign and malignant diagnosis for pulmonary nodules: extracting a variety of imaging features from true nodules and the text features of the patient which include clinical features and serum tumor markers.The features were analyzed by receiver operating characteristic(ROC)curve and hypothesis test method,and the results showed that serum tumor markers had a higher ability to discriminate benign and malignant pulmonary nodules.In this study,the fusion feature of image feature and text feature is used as the input of support vector machine(SVM),and the fusion characteristics are compared with the classification prediction results of individual type features.The results show that the accuracy of benign and malignant diagnosis of pulmonary nodules can be improved effectively by the fusion feature.By comparing the diagnosis effect among K-nearest neighbor(KNN),random forest(RF),SVM,low qualification doctors and high qualification doctors,it was found that most of the parameters of the classifiers were between high and low qualification doctors,SVM classifier has a relatively good diagnostic effect.The experimental results show that the fusion features including imaging features,clinical features and serum tumor markers can improve the diagnostic effect of benign and malignant pulmonary nodules,so as to assist the doctor better.