Research On Algorithm Of CT Image Segmentation Based On Graph Cuts

Medical image segmentation is the technology and process to separate an image into certain regions with special characteristics and to extract the interesting target. It's one of the key issues in medical image processing, visualization and computer-aided diagnosis. And the segmentation result can directly affect the accuracy of subsequent processes. So the research is the hot spot all the time.Medical image segmentation based on graph cuts is a new developing method recently. Graph cuts algorithm is the technology to convert an image into s-t network constructed in the way of Markov Random Field and optimize it using max-flow /min-cut algorithm. Its core idea is to construct a reasonable energy function, and minimize the function by combinatorial optimization technique. The novelties of the algorithm lie in its combination of regional and boundary information, global optimization, non-topological restrictions, combination with a variety of priori knowledge and so on.In the paper it introduces the background of medical image segmentation, including the history of medical image segmentation, segmentation evaluation criteria and medical image algorithm platform, and then presents the basic knowledge of graph cuts algorithm, including the basic knowledge of graph theory associated with graph cuts algorithm, max-flow/min-cut algorithm and the basic procedures of graph cuts. All the results of the previous studies constitute a good basis for this paper.Supported by graph cuts theory, it makes an attempt to research graph cuts algorithm from the efficiency, accuracy and interactive ways:(1) Use mean shift algorithm as a preprocess step of graph cuts algorithm to improve its efficiency. The idea is as follows:process the original image with mean shift method, and make use of the over-segmentation regions as a whole named as "super-pixel" instead of individual pixel to run graph cuts algorithm, which can greatly reduce the actual number of vertices and edges in the graph and significantly improve the efficiency of graph cuts algorithm. Mean shift, watershed method and clustering algorithm are all super-pixel methods. But watershed algorithm is sensitive to noise and has a bad ability of remaining boundary, and in the clustering algorithm it must set the number of clusters. Mean shift algorithm has good properties of remaining boundary, inhibition of noise and no need to set the number of clusters. So it chooses mean shift method as a super-pixel method.3D vertebra CT images are tested for the algorithm. The average symmetric surface distance by this algorithm is about 1.208mm, while the same index by watershed and graph cuts algorithm is about 2.924mm, the surface rendering results can also prove this method is efficient.(2) Probabilistic atlas and graph cuts are combined to improve the accuracy of graph cuts algorithm. To obtain more accurate result, priori shape knowledge is added to the framework of graph cuts method. In the paper, it chooses probabilistic atlas on behalf of shape information. It's a non-parameter shape model and can simplify the complexity of modeling. All the training images are mapped to the image to segment, after some operations, it can get an image with the same size and gray value stood for the probability that the pixel belongs to certain organ, that is probability atlas. In the paper, the probabilistic atlas, as a priori shape, and the original image are combined as the regional information index, and so the high level information can instruct the segmentation process. It segments the liver from the abdominal CT images to test the algorithm. Manual segmentation result is set to 75 point and this study is about 69 point. It can basically reach clinical requirements.(3) Turn interactive segmentation to automatic segmentation. For a 3D medical image, the experienced radiologist trains the data and obtains the mean and stood deviation of foreground and background as the priori information of the anatomical structure. The information is put into the codes and can stand for most of this kind of medical images contained the anatomical structure. All the graph cuts algorithms in the paper are automatic segmentation in this way.