Study On Image Segmentation Of CT Image Sequences Based On Fuzzy Model And Shape Feature

Computed tomography(CT) technology,an advanced detection technology, can directly and clearly obtain the internal structures and properties of measured objects in the form of images. It is widely used in the area of medical diagnosis and industrial non-destructive testing. With the development of CT technology and the improvement of application requirements in CT, quantitative and automatic analysis and measurement in image processing technology of CT system, to overcome the deficiencies in subjective qualitative evaluation, is an important aspect of the development of CT technology. Segmentation on CT images is the key technology and difficulty to realize quantitative analysis and automatic recognize and measurement of CT images. Segmentation on CT image sequences is considered as the research content in this paper. In response to the issues of image segmentation on one type of fuzzy medical CT image and one type of industrial CT image with anisotropic voxel size, this paper proposes an object membership based automatic anatomy segmentation method for medical whole body positron emission tomography/computed tomography(PET/CT) images, and a shape feature based crack segmentation method for industrial CT images. The main contents are as follows:1. An object membership based automatic anatomy segmentation method for PET/CT images is proposed. To solve the problem that although automatic anatomy recognition(AAR) method can achieve segmentation result with high precision in anatomy segmentation of diagnostic CT images, it obtains segmentation result with low precision in anatomy segmentation of PET and low-dose CT images, this paper proposes an object membership based automatic anatomy segmentation method with the use of characteristics of gray values and texture properties. Firstly, in the process of model building, an object membership function which combines gray values and texture properties of training images is proposed, with the aim of calculating the possibilities of existence of objects. Then, in the process of segmentation, the object membership function is used to obtain object membership values of the test image, then the object model pose is initially determined and refined. Finally, the optimal pose of object model is obtained, and the space distribution of measured object is achieved. The experiments is conducted with PET/CT images and two metrics including location error and scale error. The experiments demonstrate that our method can achieve anatomy segmentation results with higher precision when compared with results of the original AAR method. In the results of our method, average location error is as low as 1-2 voxels and average scale error is up to standard value 1.2. The optimal threshold training method of AAR is improved. To solve the problem of high dimension of search space and poor adaptability that it is only suitable for original gray images, in the original optimal threshold training method of AAR, this paper proposes a new optimal threshold training method with the use of super masks and cumulative histograms. Firstly cumulative histograms of object and background are separately computed in a super mask, then the absolute difference between the area of the cumulative histogram of object and the cumulative histogram of background defined by any possible threshold is calculated, and the optimal threshold is selected by a criterion, that the threshold achieves the maximum absolute difference value. The search space of the improved optimal threshold training method is of one dimension, which ensures high processing efficiency and avoid omitting any possible thresholds caused by the limit of search space of threshold intervals. The experiments demonstrate that the improved method can be applied in intensity, texture and membership image to obtain the appropriate thresholds of objects, leading to anatomy segmentation with higher precision.3. The hierarchy of AAR method is improved. The original AAR method is only applicable to body region images, such as thoracic and abdominal images, and requires manually trim from whole body images to body region images. To improve the level of automation, whole body hierarchy is proposed based on anatomy relationships between organs in whole body, which means all the organs in the body is organized in a hierarchy with the representation as a tree and the model building and segmentation steps of all organs proceed hierarchically following the hierarchy in a breadth-first manner. With whole body torso PET/CT images, the experiments demonstrate that the improved method can achieve segmentation results of whole body torso anatomy with high precision and improve the level of automation.4. Inter-modality model building-initial segmentation strategies are proposed. In most model building methods, intra-modality model building-initial segmentation needs training data from the same modality, and ignore the general rapid prototype of various modalities. In this paper, based on two main steps of AAR, model building and initial segmentation, inter-modality model building-initial segmentation strategies are proposed, which takes advantages of modality-independent of fuzzy model and shape and location information in fuzzy model. The experiments demonstrate that the models built from diagnostic CT data can be deployed on initial object segmentation of low-dose CT, PET and their object membership images, suggesting the potential of rapid prototyping for various imaging modality.5. A shape feature based crack segmentation method for industrial CT image sequences is proposed. The detection, automatic visualization and measurement of cracks in work-pieces is one of the difficulties to be settled in industrial CT, and image segmentation is the key point. In industrial CT systems, the acquired three dimensional images consist of sequences of image slices. On one hand, the size of voxels in the direction of slice plane is different from that in the direction perpendicular to the slice plane, and in some cases this difference is even higher than 10 times. On the other hand, the artifacts in industrial CT images are heavy, can adversely affect crack segmentation and measurement. In order to solve these problems, a crack segmentation method of anisotropic industrial CT image sequences is proposed. Firstly, we adopt Hessian matrix based two dimensional linear structure filtering to enhance linear regions of the image. Subsequently, we propose a novel two dimensional histogram incorporating the inter-layer continuity of gray values and directions and the inner-layer average values of linear neighborhoods in the filtering image, which weaken the adverse effect from the artifacts. Then the threshold is determined by the maximum class entropy of the histogram. Finally, the binary segmentation results for cracks are achieved. The experiments is based on industrial CT image sequence of real work-pieces. Three metrics are used for evaluation, including accuracy, recall and F1 values. The experiments demonstrate that the proposed crack segmentation method not only achieves more complete and accurate segmentation results, when compared with four other methods, meeting the requirements of accuracy in real applications of crack segmentation on industrial CT images, but also improves the level of automation.