Pulmonary Nodule Feature Extraction And Aided Detection Based On Medical Image

In recent years, many countries’ statistics have shown that the incidence and mortality of lung cancer continue to increase, now have ranked first of all malignancies. Pulmonary nodules are one of the most important early signs of lung cancer. According to the lesion characteristics of pulmonary nodules, we can infer the lesion character of lung lesions, and help prognosis. Therefore, timely lung nodule detection and treatment to patients are key measures to reduce lung cancer mortality. The best way to diagnose lung cancer is CT tomography scanning, but to screen out the lung nodule from a large number of CT images is a heavy work for the doctors, meanwhile, the work is subjectivity, easily lead to misdiagnosis and missed diagnosis. Currently, to conduct initial lung nodule detection and classification of the patient’s lung CT image through computer-aided diagnosis technology, which can provide useful reference information to doctors, help doctors make an accurate diagnosis of lung disease patient, therefore, the research work of computer-aided diagnosis technology become more and more important, this is what our paper’s main purpose. This paper’s main research work is to carry out two important aspects of lung cancer computer-aided diagnosis, including feature extraction and classification.The main workflow of lung CAD system is: segmentation of lung nodules, feature extraction of lung nodules and classification of pulmonary nodules, the main technology used in which involves digital image processing and machine learning. Because pulmonary nodule segmentation technology has relatively mature, this paper use the existing high accuracy FRFCM(Fast Anti-noise FCM) clustering segmentation method to conduct lung CT image segmentation. The main work of this paper includes the following two aspects:1.In feature extraction part, this paper conduct a comprehensive analysis and study of the pulmonary nodules morphology, density, boundary, location and spatial information of five areas’ medical pathology and image information. Based on the medical experts’ labeling information of pulmonary nodules in LIDC(Lung Image Database Consortium), we proposed a set of mainly shape-based feature vectors to fully characterize the pulmonary nodules.2.Classification of pulmonary nodules is a key link in lung cancer CAD system. In this paper, we propose a modified partial supervision fuzzy clustering algorithm after studying and analyzing the limitation of the existing classification method. Besides the introduction of labeled samples, we calculate the reference membership by exploiting the class information of labeled samples in the process of clustering, and use the reference membership to guide the clustering process of the testing samples, for helping the testing samples to cluster more accurate.The results from the classification of pulmonary nodules shows that the improved algorithm proposed in this paper finally achieve better results than the common methods. The average classification accuracy rate reached 77.6 percent, where the malignant nodules recognition accuracy rate reached 91.3percent.