Research On Pulmonary CT Image Analysis And Feature Extraction

Lung cancer is one of the most deadly diseases in the world, it threatens people's lives. Detection and therapy in the early stage is the most effective way to prevent and cure the pulmonary diseases. Most pulmonary diseases usually come up as Solitary Pulmonary Nodules (SPNs) in CT images. So it is worthwhile to apply computer-aided diagnosis techniques for the early detection and recognition of pulmonary diseases.Computer-Aided Diagnosis (CAD) techniques provide potential necessaries for the early detection and recognition of pulmonary diseases. By using the digital image processing and pattern recognition techniques, a CAD system has advantages for medical imaging assessment. On the one hand, it can help physicians to reduce the heave workload and improve the working efficiency. On the other hand, it can provide more objective information for physicians to make further diagnosis.Due to the large amount of CT images, complicated lung structures, calcification, nodule density, shape, texture, surrounding structure etc., it is necessary to provide physicians with quantitative measurements of features in order to make further diagnosis. And feature extraction is one of its most important phases.Focusing on pulmonary nodules feature extraction, a general scheme for nodules feature extraction is proposed in this dissertation after analyzing and combining the medical signs of nodules in CT images and the expert knowledge. This scheme is analyzed and realized from five aspects, including gray level, shape, texture, spatial context and outer features, to quantitate the regions of interest (ROI) in CT images. And then, an assessment is designed to evaluate the effectiveness of the features extracted by using the scheme mentioned above. And the experiment results indicate the utility of the feature extraction scheme. The performance of nodules detection reaches 98.5%.This dissertation is aiming to provide systematic and quantitative measurements on pulmonary nodules in CT images.First, the medical signs and relative knowledge of pulmonary nodules are summarized. The medical signs of nodules are named according to their complicated appearances and have their special meanings. And different signs indicate different state to diagnosis. So it is necessary to analyze the important medical signs.Second, a scheme for shape features extraction is investigated. The important features, such as lobulation, burr and calcify, are quantitated by using edge detection and D-P (Douglas-Peucker) algorithm to reduce the image data from 2-D into 1-D key points in the curve. In addition, the shape of fractal theory is applied to calculate the whole shape of the ROI.Third, a scheme for texture features is designed. The texture features of the nodules are described based on wavelet coefficients Gaussian density distribution (GDD). Besides this, a method by calculating the region variance of key points on the edge is proposed to quantitate the clarity of the edge.Fourth, a scheme for spatial features is presented. The position information of the ROI in the local CT slice and the similarity of the relative region in the adjacent slices are calculated to describe the spatial context features.Fifth, an assessment is designed to evaluate the effectiveness of the features extracted by the above methods. On the one hand, ROC (Receiver Operation Characteristic) is used to evaluate the relativity between features and medical signs of the nodules. On the other hand, a new feature measurement RMI (Ratio of Mutual Information) is presented based on the concept of rough set theory. Then a novel heuristic algorithm, named MRMI-UC (Algorithm based on Maximal Ratio of RMI and Uncertainty Coefficient), is proposed for Feature Selection based on rough set theory. And these two are used to estimate the relativity between features and property (benign or malignant) of the nodules.