Analysis And Identification Of Pulmonary Nodule Images

Lung cancer ranks as the leading cause of cancer death among men and women, the incidence of the disease stands at the second. For the early performance of lung cancer is pulmonary nodules, the detection and diagnosis of pulmonary nodule has the importance of early diagnosis and treatment of lung cancer. In addition, metastases are prone to occur in lung, and they are common malignant lung nodules. So Identification of metastases and benign nodules is of clinical significance for further treatment. CT plays an important role in pulmonary nodule detection and characterization. In recent years, with the application of multi-slice CT (MSCT), high-resolution CT (HRCT) and low-dose CT (LDCT), the detection rate of small pulmonary nodule and early lung cancer is rising. In order to reduce the possibility of omission or mistake in pulmonary nodule detection, and to improve the detection rate of nodules as well as diagnostic accuracy, computer aided image processing technology must be applied for automatic pulmonary nodule image analysis and recognition, to assist the evaluation and diagnosis of radiology physician.The key technologies realizing automatic analysis and recognition of pulmonary nodules are: a series of research and application algorithms of image processing, analysis and understanding on pulmonary nodule segmentation, detection, diagnosis, etc. Because the characteristics of medical images are: fuzzy edged, non-uniformed gray scale, of individual differences, and with artifact and noise, the relevant algorithms to achieve the sensitivity and accuracy requirements are on higher difficulty. In this paper, applying to the medical characteristics of lung CT image, we conducted an in-depth research and a large number of experimental on lung segmentation, suspected nodule analysis and extraction, and pulmonary nodule detection and diagnosis.In this paper, the main contributions are detailed as follows:(1) Fast lung CT image segmentation.In order to improve the accuracy of detection and diagnosis, most of the lung CAD methods do image pre-processing first, thus, determine the borders of lungs. This step is called "lung segmentation". In the lung segmentation process, because of the medial area before the mediastinal between the left and right lung is relatively narrow, the region and the lung area is often very low contrast caused by the existence of partial volume effect phenomenon, and the segmentation of left and right lung will failed.This paper put forward the base frame of lung segmentation first. Using optimal thresholding and mathematical morphologic methods, the rough image of lung is acquired. Then, we present a fast self-fit segmentation refinement algorithm, adapt to the unsuccessful left and right segmentation of thredsholding.(2) Self-fit morphologic pulmonary nodules segmentation.Automatic Segmentation of pulmonary nodules is the premise of accurate extraction of the pulmonary nodules characteristics. The main segment features of pulmonary nodules are: the CT value of the nodular center is high, and the nodule borders are irregular curve closed or semi-closed. Because of the lack of golden standard (ground truth), related comparison and evaluation of pulmonary nodule segmentation methods are very difficult. The problems of most of these methods existed in: The model should not apply to all types of nodules, particularly the nodules neighboring blood vessels, trachea, pleural, or irregularly shaped nodules, or nodules with "halo sign".This paper put forward the automatic process of pulmonary nodule segmentation firstly. Then, using region growing method based on the contrast and the gradient, we acquire the pulmonary nodule image. Finally, we present a self-fit morphologic segmentation algorithm, adapting to the unsuccessful nodule segmentation of region growing.(3) Automatic pulmonary nodule detection and classification.One of the conventional pulmonary nodule evaluation methods are: to make a comprehensive analysis of its shape, size, density, enhanced mode and characteristics of growth patterns. For different types of nodules, the recognition accuracy of single pulmonary nodules is relatively high, but the detection rate of other types of pulmonary nodules is not high because of various anatomical structures and pathological factors, especially the nodules neighboring blood vessels, trachea, pleural, or nodules too small, or nodules with "halo sign".In this paper, automatic detection of pulmonary nodules is achieved as follows. First of all, analyze and extract basic pulmonary nodules' gray-level, shape and spatial characteristics, such as: the mean and standard deviation of gray-level, area, radius, similarity degree of circularity, etc. Do associated measurement and calculation. Then, apply Bayes optimal statistical classifier and linear discriminant Analysis (LDA) classifier respectively on a preliminary classification of suspected nodules. After that, in order to improve the accuracy of pulmonary nodule detection, use radial gradient index (RGI) characteristics to remove false positives (FP).(4) Automatic identification of benign and malignant pulmonary nodules.To experienced physicians, the identification and diagnosis of pulmonary nodules are problems, let alone computer systems. To improve the identification methods and the diagnostic accuracy are very difficult. At present, the study for diagnosis of benign and malignant pulmonary nodules mainly concentrated in a variety of machine learning methods. Because the feature Data of pulmonary nodules is high-dimensional and nonlinear, these methods give many pressing problem such as high dimension disaster, failure of linear model, etc.This paper studies Support Vector Machine SVM classification as the solution to the limited study sample set, using SVM with slack variables and penalty factors as classification method of benign and malignant pulmonary nodules.Although this paper achieved a certain degree of progress on improving lung and lung nodule segmentation accuracy and reducing false-positive diagnosis of pulmonary nodule, there are still a number of shortcomings:(1)The collection and collation of pulmonary nodules CT images is not so norms, now the data are failed to form a large complete database of pulmonary nodules with enough scale and nodular type. (2) Only the segmentation of solitary pulmonary nodules and nodules associated vascular or pleural have been studied, while deal with other irregular nodules have to be the next phase of work; (3) The characteristics of pulmonary nodules is also limited to the present characteristics of the existing literature, and the LDA classifier training has not yet carried out enough experiments, so nodule detection rate have to be further improved; (4) The research and experiment of identification of the benign and malignant pulmonary nodules only conducted a preliminary, further refinement should be carried on the basis of theoretical and the practical application.