Study On Accurate Segmentation Algorithms 3d Medical Images

As the development of computer techniques and medical imaging modalities, medical images can not only show the 2D projection of the patient body, but also the 3D volume. Meanwhile, the resolution of the imaging is becoming higher and higher. Medical image segmentation, as an important preprocessing step for many image based medical applications, such as computer aided diagnosis and computer aided treatment, is a technique that separates the medical image into several regions; and each region is satisfied by a predefined criteria. Unfortunately, most of the previous research on the medical image segmentation focused themselves on the 2D images. Even these method are applied on the 3D images, they usually use a slice by slice manner. One major weak point for the slice by slice techniques is that they cannot incorporate the information along the Z direction. However, applying the algorithms on the 3D medical volume is difficult due to the following reasons: (1) It is difficult to extent the curve on 2D image to surface on the 3D volume, due to the limitation of some techniques. For instance, there is no counterpart in 3D (surface) of the curves in 2D for the shortest distance path algorithm in graph theory. (2) Some 3D segmentation methods are so slow that are not suitable to use in practice. (3) It is not convenient to interactive with the volume due to the limitation of the visualization technique.After reviewing many popular segmentations methods, we proposed several 3D automatic and interactive segmentation schemes. Specifically, the main contributions of this work are as follows:1. Proposed a weighted center median based bilateral fitter, which is a combination of the median filter and bilateral filter. The proposed filter is able to surpass the impulse noise and Gaussian noise simultaneously.2. Proposed a hybrid scheme which can extract brain tissue from head MRI T1 scans. Given the rough segmentation by graph cut method, marching cubes based isosruface extraction is performed to extract the rough brain surface, which evolved to the exact brain surface governed by the parameterized deformable model. 3. Proposed embedded edge confidence based live-wire segmentation method. This interactive method considered the weak but sharp edges, which are common presented in medical images. Segmentation results indicated that the proposed method required less control points compared with the conventional live-wire method, which means increased working efficiency of the operator.4. Proposed nonlinear kernel based interactive segmentation method using cellular automata. The proposed method increased the convergence speed and reduced the status switch time for the cells on the boundary compared with the classical method.5. By analyzing the architecture and generous propose computing method of the GPU, we proposed a GPU implementation of the 3D cellular automata based interactive segmentation method. Experimental results shown the 90 times faster compared with the CPU implementation.Finally, the conclusion of this dissertation and the prospect of the research are given.