Segmentation Of Brain MR Images Based On Non-local FCM

With the aggravation of the degree of the aged in China, the incidence of brain diseases is increasing day by day. Therefore, it is very important to diagnose the disease with the aid of medical imaging technology. Magnetic resonance imaging (MRI) because of its on the human body without any ionizing radiation injury, soft tissue had higher resolution, multi imaging parameters, contains a large amount of information and other advantages, has been widely used in medical image diagnosis.Fuzzy c-means (FCM) algorithm is a classical clustering method and it has advantages such as unsupervised study, simple realization, fast operation, etc. However, in fact, most of the brain MR images are more or less have noise and bias which make it difficult to use the traditional FCM algorithm to get the results of the ideal segmentation. Therefore, in this paper, we want to introduce the spatial information of images to improve the objective function of traditional algorithm, and integrate the bias correction into the framework, so that it can reduce the effect of noise, and recover the bias field. The main research work of this paper are as follows:(1) To reduce the impact of noise and bias field, an image segmentation and bias correction model based on improved FCM with non-local spatial information is proposed. This model takes spatial structural similarity measurement between the image patches into consideration, Meanwhile the bias field is taken into the model, thus the proposed model can reduce the bias and noise effectively as well as keep more structural information.(2) Aiming at the problem of robustness to the traditional model and the segmentation accuracy is not high enough, a Fuzzy Clustering model based on non-local spatial information is proposed. Firstly, integrating the non-local information into the model can reduce the impact of noise as well as keep the image structures. Secondly, constructing the image gray scale distribution by the multivariate Gaussian distribution and set it as the distance function to reduce the lack of robustness caused by the traditional Euclidean distance. Finally, to overcome the impact of intensity inhomogeneity, the bias field is approximated at the pixel-by-pixel level by using a linear combination of basis functions, and parameterized by the coefficients of the basis functions.