Research On Algorithms Of Lesions Segmentation And Classification For Pulmonary CAD

Lung cancer is one of the most common malignant visceral tumors, and it is usually diagnosed in later period with lower survival rate, so early diagnosis is significant. At present, technology of CAD in medical imaging emerges continuously and develops rapidly, and has become a hot issue in this field. Pulmonary CAD system can help doctors assess the medical images, improve diagnostic efficiency and reduce the burden on doctors. Studies show that CAD has played a positive role in early detection, early diagnosis, improving diagnosis accuracy and reducing misdiagnosis of lung cancer.In this thesis, algorithms of lesions segmentation and classification for pulmonary CAD are researched, the main work includes the following three parts:(1) Represent the initial lung contour. In order to analysis lesions better, for the conglutinate plumonary nodule, the lung contour is represented. First, the common method, rolling sphere method (that is, closed operation in mathematical morphology), is adapted to work, and experimental result is showed, the shortage that the size and shape of the structural element are unxertain is discussed. Then, for each point on the lung contour, using the chain code to classify them as a concave ponit, or a convex point. Using the bresenham algorithm is to represent the lung contour. Experiments prove that the represent result is very well, and the shortage of "rolling sphere method" is balanced.(2) Research on the segment of region of interent (ROI). For the pulmonary parenchyma after segmentation, the Otsu algorithm is used to segment pulmonary parenchyma initially, and the region of interest is gotten. Then through analysis differences from nodules and non-nodules, surface normal overlap method is improved, adaptive surface normal overlap is made. Experiments prove that potential nodules objects are segmented so well.(3) Designing classifier to class nodules and non-nodules. Because of unbalanced data between nodules and non-nodules, the classifer to balance datas based on support vector machine (SVM) is researched. SMOTE is to make synthetic minority datasets, biased_SVM is to solve the problem of offset. The confusion matrix and functions of F-measure and G-mean is used to assess classifiers, and the inaccuracy of accuracy rate is analyzed.In this thesis, the main idea about study is using some simple examples to verify the feasibility of algorithms at first. And then the correct algorithms in the Pulmonary CAD are uesd. This idea can ensure the reliability of algorithms.