Segmentation And Centerline Extraction Of Tubular Structures Based On CT Images

In the process of computer-aided diagnosis, the segmentation and centerline extraction of human tubular structures based on CT images are two essential steps in solving several practical applications such as quantitative analysis of diseases, visualization and registration of patient images obtained at different times. However, the tubular structures in human being and their background are usually very complicated. For this reason, the segmentation and centerline extraction of tubular structures from CT images are two especially challenging problems.According to the practical problems, the research direction of this dissertation is towards developing faster, more accurate and more automated techniques of the segmentation and centerline extraction of tubular structures. The main finished works in this dissertation include:First, a fast and automatic ascending aorta segmentation method is proposed. The method applies the two-dimensional region growing algorithm to segment ascending aorta slice by slice in an iterative procedure without any interaction. The proposed method has been evaluated on103real cardiac CT angiographic datasets. The evaluation results show that the method can segment ascending aorta fast and robustly. The success rate of the method is95.1%and the average running time is1.5s.Second, an automatic coronary segmentation method based on CT angiographic images is proposed. The method applies tubular structure similarity function to extract coronary seed points automatically in scale space. Then according to the structure features of coronary artery, an improved layer region growing algorithm is used to segment coronary artery. The evaluation experiments show that the method can segment coronary artery accurately and fast.Third, three coronary centerline extraction methods are proposed, and they are the improved tracking-based method, the thinning and tracking-based method and the improved distance transformation-based method, respectively. Then, the accuracy, running time, automaticity and application range of the three methods are compared. Moreover, three automatic centerline correction methods are proposed to correct centerline errors arising from narrow-neck over-segmentation, wide-neck over-segmentation/partial under-segmentation and total under-segmentation, respectively. And the quantitative evaluation results show that the proposed methods can effectively correct centerline errors and improve the accuracy of centerlines of coronary arteries.Fourth, an improved cerebral and carotid artery segmentation method based on registration/subtraction and morphologic operation is proposed. The experiments show that the improved segmentation method can segment cerebral and carotid arteries accurately and completely even when the quality of the images is poor. Moreover, two vessel edit methods are proposed to refine the segmentation results of cerebral and carotid arteries. They are used to remove wrongly segmented tissues from segmentation results and add missing vessels to segmentation results, respectively. The quantitative evaluation results prove the high accuracy and repeatability of the proposed methods.Fifth, a robust and fast method is proposed to extract inferior alveolar nerve (IAN) canals from CT images. First of all, the method identifies the center points of an IAN canal on cross-sectional images using feature analysis of connected regions. Then, the centerline of this IAN canal is computed by using a spline interpolation. Finally, the IAN canal is extracted by adding the voxels with similar characteristics to the centerline region. The evaluation results show that the proposed method can successfully extract all the IAN canals in the evaluation datasets (100%success rate) and both the accuracy and the processing time are satisfactory.