Research On The MRI Brain Image Segmentation Algorithm

With the clinical application of all kinds of imaging equipments, the medical image post-processing technology has become the hot issue in the recent image segmentation field, while the image segmentation is the premise and foundation of the three-dimensional reconstruction, visualization, registration, integration and quantitative analysis techniques. Based on the research of several usual image segmentation algorithm, the paper focuses on the use of Graph Cut algorithm in the MRI brain image. And the task is to achieve the partition of brain gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF).For the minimum cut solution of the problem, this paper adopts the "preflow" algorithm for the maximum flow of the graph network in order to attain the minimum cut set. The method achieves the image segmentation algorithm and improves computational speed.Owing to the normalized cut (Ncut) algorithm's weakness of calculating slowly, the paper uses the watershed transformation algorithm in the graph cut framework and propose a new speedy normalized cut algorithm—sNcut algorithm. The graph cut is applied to the image segmentation, whose core is to map an image into a weighted undirected graph based on a certain model and then search for the optimal cut set in order to achieve the partition of the object region and background region. A number of small regions, split by the watershed transformation, whose internal pixels are closely related to one another, are mapped the vertices in the graph. This method can significantly reduce the number of vertices and the computing speed has improved significantly than the direct pixel mapping method.Direct at the over-segmentation problem in the watershed processing and the phenomenon that noise signal encourages the over-segmentation, the paper designs a multi-structure adaptive morphological filters to remove the image noise and proceeds the region combination by adding marker points. The pre-processing and post-processing of segmentation effectively inhibits the phenomenon of over-segmentation to ensure that the result of the pre-segmentation is meaningful and to provide a prerequisite for the graph construction and searching for the optimal cut set.In addition, in order to ensure the accuracy of the normalized cut segmentation, the paper improves the method of determining weights. The general approach is to use the pixels'gray gradient to determine the weight and establish the graph model. The paper combines the pixels'grayscale and space information to determine the weight between the vertices and proposes a new model of weight--indensity and spatial information (ISI).In order to verify the effectiveness of the algorithm, this paper uses 40 sets of brain image data for simulation experiments. However the segmentation results of 35 sets of data achieves the intended purpose and cerebral gray matter, white matter and cerebrospinal fluid and other organizations can be partitioned accurately. The other five sets of the data can not be divided or split error due to image quality. The correct rate of segmentation is87.5%, and the experimental results show that the proposed algorithm is effective.