Research And Application Of Medical Image Segmentation Methods Based On Deformation Model

Although deformable model was originally developed for application to problems in computer vision and computer graphics, its potential for use in medical image analysis has been quickly realized. This thesis focused on investigating the theoretical background of deformable models and improving them for applications in segmentation of medical images for computer-assisted surgery navigation system.There are basically two types of deformable models: parametric deformable model and geometric deformable model. Parametric deformable model, which allows direct interaction with the model and leads to a compact representation for fast real-time implementation, expresses curves and surfaces explicitly in their parametric forms during deformation. On contrast, Geometric deformable model implicitly represents curves and surfaces as a level set function. Furthermore, it can handle topological changes naturally and therefore rather fits 3D image segmentation.The energy minimizing formulation and dynamic force formulation of parametric deformable model were described respectively. Five typical kinds of external forces for parametric deformable model (traditional Gaussian potential force, multi-scale Gaussian potential force, pressure force, distance potential force, gradient vector flow model) as well as several numerical methods to implement deformable models (gradient descent method, dynamic programming, greedy algorithm, discrete dynamic contour model) were listed, discussed and compared respectively. Among them, the discrete dynamic contour model(DDCM) was a fast and stable method. Then an improved DDCM integrating the ideal of Gaussian potential force and pressure force was proposed, which was more robust, adaptable and efficient.Several problems in the application of Snakes, including resampling, parameter selection as well as curve intersection, were discussed and their solutions were proposed.Level set methods were discussed in detail, which is the basic of geometric deformable models. Several numerical methods were introduced to implement the Level set method and their qualities were compared, such as Narrow Band method, Fast Marching method and Group Marching method.The speed function defined in Level set method can be divided into two classes: function based on the gradient of image and function based on the image gray. The former defines the speed function F based on the gradient of image, while the F is sensitive to noise and may leads to wrong results at the indistinct or disconnected boundaries of structure. Chan and Vese proposed a new method based on Mumford-Shah image segmentation model, which was based on global image gray information and can deal with the fault of the former method. In this paper, the 2D Chan-Vese model was extended to 3D space and the numerical implement method was improved. In addition, the 3D Chan-Vese model was employed to extract brain tumors. A lot of experiments showed that the improved 3D Chan-Vese model was efficient and valuable in clinic application systems.Finally, the improved methods in this paper were applied to the computer-assisted surgery navigation system. The improved DDCM method and the improved 3D Chan-Vese model were used to extract the head surface and brain tumors respectively. Relatively satisfying results of the segmentation were attained.