Research On Interactive Segmentation Of CBCT Three-dimensional Images

Recent years, Cone Beam Computed Tomography (CBCT), as an advanced technique of high resolution three-dimensional imaging, has greatly expanded human ability of visual information acquisition, and has been widely used in medicine, archaeology, industrial non-destructive testing, reverse engineering and other fields. In practical applications of CBCT, quick and precise segmentation of objects of interest is essential for subsequent measurement, analysis, reconstruction and rendering. As one of the key factors affecting the efficiency of the three-dimensional imaging system, image segmentation has become a research hotspot.Three-dimensional images usually contain many kinds of noise and artifacts, which are generated by CBCT imaging system. Automatic image segmentation algorithms can only achieve good effects in certain conditions, and have a poor adaptability and a high complexity, so that lack the ability to meet practical needs; Interactive segmentation methods introduce users’ understanding of target areas into the segmentation, which provides a guide for the image segmentation process to obtain more precise and effective results. Current researches of three-dimensional interactive image segmentation mainly focus on how to make full use of users’ feedback information based on the features of three-dimensional image data, and to improve the real-time and accuracy of interactive segmentation methods. This thesis analyzed the properties of the three-dimensional image data generated by CBCT, aiming at solving the problems such as the large amount of data, noise interference andcomplex structures of image targets, based on the three-dimensional target image denoising, combining the texture information and the similarity of local characteristics of targets in the three-dimensional image, studied the CBCT three-dimensional interactive image segmentation method and designed a new way of user interaction. The main researches and innovations are as follows:1. For suppressing the influence of three-dimensional image noise on image segmentation, a new processing method based on volume morphology is presented. While suppressing the noise, this method could keep the structures in detail of three-dimensional targets. Taking the structure distributions of image targets in three-dimensional space into consideration, this method builds two basical nonlinear operators as volume open and volume close. These two operators combine both the information gray level and the volume of the connected region, and through the decomposition and composition of volume data; the noise suppression problem is simplified into the problem of removing connected regions in binary volume data. They could smooth glitches and fill holes in three-dimensional regions while keeping the shapes and details of regions of interest, and provide good data foundation for subsequent image segmentation. In addition, a gray level decomposition method based on BitPlane and a connected region reusing method are both proposed to overcome the efficiency bottleneck steps. They could improve the computation efficiency of volume morphology without affecting the accuracy.2. For the inefficiency of traditional GrowCut interactive image segmentation method, which updates the states voxel by voxel, a fast three-dimensional interactive segmentation method based on supervoxel texture features is proposed by combining three-dimensional texture and local gray level distribution of CT Images. This method pre-segments the volume data by SLIC method, and introduces the information of the distributions of both cellular texture and gray level global distribution, and adjusts the similarity measurement between cellular dynamically. It can improve the efficiency of cellular automata iteration calculation and restrain the segmentation results in local optimum situation. The experimental results show that the method can improve the speed of interactive segmentation, and obtain reliable segmentation results.3. For the huge amount of three-dimensional image data and the complex user interaction, a three-dimensional image interactive segmentation method base on local similar features is proposed. It uses the similarities of local feature in three-dimensional space and users’feedback. In the three-dimensional space, the local features of user labeled image regions should be similar with the local features of regions containing the labeled ones. This method describes local structure information using BRISK descriptor and measures the similarity between BRISK descriptor images of adjacent slices. BRISK-Flow method is used to estimate the displacement field. The limited labeled regions are extended automatically to regions with similar features in three-dimensional space, and then the labeled regions are used to image segmentation. Experimental results show that this method can reduce the user labeling work effectively, and segment the interested regions precisely.4. For the long response time, complex interaction and the difficulties in grasping the global interactive information for users, a new user interface for the preview of segmentation results and the statistical function of labeled region information is designed. It can provide references for the interaction to rapidly modify the labeled regions and improve the segmentation accuracy. A new extendable three-dimensional image interactive segmentation platform is designed and implemented. It incorporates the proposed noise suppressing and image segmentation methods. And multi-thread processing scheme is adapted to improve the response ability of UI and make it more user-friendly.