3D Segmentation For KVCT Prostate Image Based On Density Peak Clustering

In recent years, the incidence of prostate cancer is gradually increasing. Prostate cancer has become an important killer of adult males around the world. The incidence of prostate cancer mainly concentrates on the middle-aged and old males. With the growth of population in our country, the incidence of prostate cancer will further increase, which may lead prostate cancer be the most serious malignant tumors of this century. Currently,radiotherapy is the most effective treatment method for prostate cancer. In the process of radiotherapy, the radiotherapy dose must be controlled accurately within the tumor region,otherwise it will cause irreversible damage on the body. To facilitate precisely positioning on tumor region, the first step in the treatment of prostate cancer is to split prostate out from the other complex tissues, then we can do further condition prediction, disease treatment planning and radiotherapy etc. Therefore, the study of prostate segmentation has extraordinary significance. Although there are many segmentation algorithms for prostate Computed Tomography(CT) images currently, most of them are two-dimensional segmentation algorithms. In this thesis, the proposed segmentation methods implement three-dimensional segmentation on a three-dimensional reconstruction image of sequence image. The main work is as follows:1. This thesis proposes a three-dimensional segmentation method of prostate CT image based on clustering on three-dimensional Superpixels. The method firstly extends two-dimensional Simple Linear Iterative Clustering(SLIC) Superpixels segmentation algorithm to three-dimensional. Then the two-dimensional feature based on gray-gradient co-occurrence matrix is also extended to three-dimensional. Then we use the features of Superpixels for Superpixels clustering for the first time to get the initial segmentation result of the prostate. Since the result is rough in the first time clustering, we carry out a second clustering on the initial results produced by the first clustering. After the second time clustering, a small part of a images still exist sticking problem, so we carry out three-dimensional morphological process to remove adhesions section and then get the final result.2. An interactive segmentation method is proposed. This method selects some points or lines as a target area by manual interaction on the base of initial segmentation results produced by clustering in the first time. Then, it mergers Superpixels around the target area according to correlation between the feature vectors of Superpixels. Finally, we obtain segmentation result of the target area.3. A segmentation amendment method based on registration is proposed. This method first selects part of prostate Kilo Voltage Computed Tomography(KVCT) images of easy to adhesion as training dataset and marks artificially the prostate region. Next, the image to be segmented is matched with the training dataset image. Then we obtain the training data image which has the largest normalized correlation coefficient with the image to be segmented. Finally, the manual annotation prostate area is mapped by coordinate transformation to the image to be segmented to achieve segmentation.