Automatic Detection Of Pulmonary Nodules And Features Analysis

Lung cancer is the most common malignant tumor and one of the lowest livability tumors after diagnosis as is known so far. It is increasing annually and now the first cause of cancer-related mortality in cities. In order to improve the survival rate of lung cancer patients, detection in the early stages has a significantly more hopeful prognosis and is the key treatment. CT scanning presents great opportunities for lung cancer diagnosis. However, the large amount of CT images caused increasing work and inevitable false diagnosis rate. Image processing techniques, for example segmentation and extraction of nodules, reconstruction and rendering of suspicious objects, make it possible that computer-aid reading images, help doctors to analyze pathological changes and other regions of interest in character and even in accurate quantity, and release the doctors' burden.The dissertation focuses on computer-aided detection of pulmonary nodules. The automatic detection of different pulmonary nodules attracts many universities and multinational corporations. We report several steps to analyze the image: first, the pre-processing of CT images; second, the extraction of pulmonary parenchyma; third, the segmentation of region of interest(ROI) and the last, features extraction and classification. The difference and the corresponding result of conventional CT and high resolution CT(HRCT) is discussed. We use a series of image processing methods such as edge detecting and components labeling to extract pulmonary parenchyma. Especially we discuss the case of nodules attached with pleural surface. Then we get ROI by using improved watershed segmentation and fuzzy c mean clustering methods. Features such as area, centroid etc are extracted and due to medical knowledge the nodules are classified based on several rules. The surface visualization of data is also discussed.Classification of malignant and benign nodules is the other main content of the dissertation. It is the first-level classification of nodules and a difficult problem upon present research. We take present methods into two categories: measurement and comparison, and herein focus on the measurement method. The measurements of geometry shape features and texture features of nodules are computed, especially we import the fractal theory to classification of pulmonary nodules, compute improved fractal dimensions of variable nodules images to describe the nodule complexity. According to prior clinical knowledge, we present algorithms to detect medical symptoms. Then linear discriminant analysis is used to classify nodules into malignant and benign ones, the discriminant scores are analyzed using Receiver Operating Characteristic Curves (ROC) method.