Research On Related Issues Of MR Brain Image Segmentation

Brain diseases, as one of the ten kinds of fatal diseases in the world, are threatening our lives and health everyday. Earlier diagnosis and treatment can rise the survival probability significantly. Magnetic Resonance Imaging(MRI), with its high accuracy, high speed, high resolution, harmlessness, is becoming the main technology used for the diagnosis of brain diseases, for this, the research on MR image segmentation has become the hot and hard spot in the field of medical image processing. The result of it affects pathology analysis and clinical therapy afterwards directly. With the characteristics of MR images considered, this thesis shows the research on the problems related with MR brain image segmentation, it is arranged as follows:(1)The preprocessing of images. First, this paper treats the images with histogram equalization, image smoothness and sharpening to get rid of the information that is useless and strengthens the useful information testability. Then, we analyze the relations between bias field and comentropy. The Legendre Polynomial Function is used to fit the bias field, meanwhile, the genetic algorithm is incorporated to resolve the function so that a precise bias field can be gotten. It can improve the accuracy of post-segmentation.(2)This artiche proposes a fuzzy MRF model through the research and analysis of FCM and MRF algorithms. This new model combines the advantages of these two algorithms. It not only uses the spatial information of the images as prior knowledge, but also threats the fuzziness of the images well. So it can segment the brain tissues with high efficiency and without supervision, also it has a good suppression to the noises.(3)Through the research on the typical C-V model, this paper proposes an improved C-V model based on level set evolvement function. The square deviation is used to establish the level set evolvement function and makes it into an ordinary differential function. So, the calculation is reduced. What’s more, a penalty function is added to the inner energy function so that it can get started automatically and has a better efficiency.(4)This paper combines the threshold segmentation with the Ncut regulations and proposes an improved Ncut algorithm. In the algorithm, the neighbour spatial information of the corresponding pixel is added to figure weights calculation and improves the robustness to noise of the algorithm. This paper uses the weighted matrix based on gray value instead of the one based on the number of pixels to describe the connection between pixels. Segmentation efficiency is improved.