Research Of Prostate Image Segmentation Based On Active Appearance Model

Prostate cancer is the common disease in senile, and also the most commonly diagnosed tumor in domestic. As the population ages and the change of domestic food structure, we can foresee prostate cancer will further increasing, there may be the highest incidence of malignant tumors in21century. Therefore, prostate cancer has become the focus of research and the hotspot. Although at present the etiology and pathogenesis of prostate cancer are still not completely understood, but the main method for prostate cancer treatment is radiation therapy. In the process of radiation treatment for prostate cancer, the first step is to distinguish the organization of the prostate from the rest in the original image effectively. In many ways, prostate image segmentation also has an important significance, such as judge the progression of treatment and computing the dose of radiation every time.In recent years, segmentation method based on model in the image analysis field has become a very effective segmentation method. This method is to match model contained shape and appearance information to the new segmentation images, which is conducted in a top-down fashion. Due to the inherent a-priori information, this approach is more stable against local image artifacts and perturbations than conventional low-level algorithms. While a single template shape is an adequate model for industrial applications where mass-produced, rigid objects need to be detected, this method is prone to be included in the model. A straight-forward approach to gather this information is to examine a number of training shapes by statistical means, leading to statistical shape models (SSM).To segment the prostate image by active appearance model, first of all, we need to build a prostate model, and then adjust the related parameters of the model through the initialization and search algorithm to match the model to the prostate image. The basic steps of constructing model are as following:shape representation, looking for corresponding relations between the shape, shape alignment, dimension reduction processing, shape model, appearance model. One of the biggest factors to the quality of the model is finding the anatomy correspondence among shapes. But because of the complexity of the3D shape structure and the quantity of the marked points in shape is very large, which causes great difficulties to finding the correspondence.From theory, different algorithm of automatic calculation of the correspondence is actually the point set registration. Researchers have proposed many methods to establish the correspondence between the point sets, such as the iterative closest point algorithm, softassign procrustes algorithm, and so on. In these methods, the shapes are represented by different number of points usually, and looking for the best similarity transformation. But the similarity transformation has the excessive constraints to the deformation of point set, which are useful to the training set with closed shape, such as bone. But this method to match the shape with large differences has poor performance even errors. Coherent point drift adopt the different ideas, which doesn’t look for similar transformation, and the result of correspondence between the point set is very good. In the algorithm of coherent point drift, space position distribution of the point set using gaussian mixture model expression, which will view the match as a maximum posteriori probability problem. The general point set registration methods are only using the space shape information of point set. However, the active appearance model not only using shape information but also using image characteristics, such as texture information. And the correspondence point in shapes in fact also has relation to some extent, relevant information may help for the point set registration. At the same time, we use space position information and image features of point set, we may make the final active appearance model more compact and more accurate. Based on this, the paper proposes an approach that combining image feature information with the algorithm of coherent point drift, using the image characteristics to adjust the gaussian mixture model, makes the points with similar image characteristics has more large probability, in order to improve the precision of matching and final segmentation accuracy of active appearance model.The new method proposed in this paper is tested on the3D prostate and liver point sets through the simulation experiments. The registration error can be reduced efficiently, the error of liver point set is decreased from1.84mm to1.54mm, and the prostate is from0.83mm to0.60mm. Moreover, the active appearance model constructed by this method can obtain fine segmentation, in3D CT prostate image. Compared with the original algorithm, the overlap ratio of voxels was improved from88.7%to90.2%.