Research On The Detection And Diagnosis Methods For Pulmonary Nodules Based On Multi-Dimensional Features From CT Images

According to the investigations of many cancer research centers and health organizations, lung cancer has been the leading cancer with the highest death rate in the world. The most efficient method for raising the living rate of lung cancer is to detect, diagnosis and therapy earlier, and the CT scan images for the chest make a possibility of the realization of this aim. Radiologists and doctors could detect and diagnosis the lung cancer directly thorough chest CT images in vision. As the rapid development of CT scanning technique, the image resolution is higher and higher, so that nodules can be detected in the reconstruction images are smaller and smaller. However, to detect the nodules only in visual is obviously very difficult in the rapid increasing image data. In addition, the gold standard of diagnosis of benign and malignant nodules in clinic is biopsy, which is an invasive diagnosis method bringing sufferings to patients. In order to assist doctors to detect and diagnosis the nodules in CT images non-invasively, computer-aided detection (CADe) and computer-aided diagnosis (CADx) systems were proposed. In this dissertation, the research work is focusing on the key and difficult points of methods in these two systems as follows.(1) This dissertation has analyzed1018chest CT image cases containing lung nodules from LIDC-IDRI deeply, which is the largest public database in the world. According to the location coordinates and important features information of the pulmonary nodules given by radiologists, the images data of nodules were extracted to be the evaluation references of CADe detection results. The image data are also used as the input data of CADx algorithms, which supported large of data for the following experiments. According to the author’s knowledge, no other research group has done the similar analysis on the whole LIDC-IDRI database.(2) For researching on the main algorithms of CADe based on CT scan images, this dissertation proposed an automatic detection and extraction method of pulmonary nodules (especially pleural nodules). First, the lung mask was extracted automatically based on the method combining of two classes VQ detection algorithm and morphology information, which was used to extract the whole lung area image automatically. Second, the initial nodule candidates were detected by using four classes VQ algorithm. Then, some empirical values were used to remove some false positive nodules preliminarily. Finally, a supervised classification algorithm based on multiple features was used to further remove the false positive nodules from the left suspicious nodule candidates. By comparison with other methods about the sensitivities, false positive rates (FPs) and speed of the pulmonary nodules detection, the results showed that the automatic detection and extraction method of pulmonary nodules proposed in this dissertation presented higher efficiency.(3) In the classification algorithm of benign and malignant pulmonary nodules based on multi-dimensional texture features in gray images, the segmentation algorithms on the surface and shape features extraction of pulmonary nodules usually affect the accuracy. To avoid the limitation, this dissertation focused on the internal structure character (namely texture feature) for the diagnosis of nodules. First, three common texture feature extraction algorithms were used on the nodules’gray images to extract the two-dimensional (2D) texture features. And by the comparisons of classification results, the performance of Haralick texture features is the best. Then, a new model of three-dimensional (3D) Haralick texture features was proposed based on the calculation principle of2D Haralick texture features, which is a deep study on the3D texture structure of nodules. Finally, based on different types of nodules, compared the classification performances of2D and3D Haralick texture features, the results showed that3D Haralick texture features have more advantage than the2D ones.(4) For researching on the validity of texture features used for the classification of benign and malignant nodules, this dissertation has proposed a new calculation method of multi-dimensional texture features from multi-level difference images (gradient and curvature images). Supposing there may be more information on the connection regions of different tissues and nodules, this dissertation suggested to study the texture features of the changes of structures in the nodule images for the malignancy diagnosis. Moreover, both the datasets including the uncertain nodules and the datasets including no uncertain nodules were studied in the experiments,2D and3D texture features were extracted respectively. Through the classification experiments on different combinations of texture features, the results showed that the combinations of the texture features from multi-level difference images and the gray image in the same dimensions can make higher performance on the classifications of benign and malignant nodules than the ones from the single type of images.