Research On Brain MR Image Segmentation With Fuzzy Clustering Based Model

Magnetic resonance imaging (MRI) has several advantages over other medical imaging modalities, including high contrast among different soft tissues, relatively high spatial resolution across the entire field of view and multi-spectral characteristics. Therefore, it has been widely used in quantitative brain imaging studies. Quantitative volumetric measurement and three-dimensional visualization of brain tissues are helpful for pathological evolution analyses, where image segmentation plays an important role. However, MR images suffer from several major artifacts, including intensity inhomogeneity, noise, partial volume (PV) effect and low contrast, which make MR segmentation remain a challenging topic. Therefore, in this thesis, we focus on brain MR image segmentation based on fuzzy clustering model from two aspects, including the construction and improvement of objective function and the uncertainty description of data. All the algorithms proposed in this thesis can get an accurate segmentation result. Our work mainly includes the following parts:(1) A novel anisotropic weighted fuzzy c-means clustering algorithm is proposed to over-come the impact of noise in the image during segmentation. Based on the discussion of the strategies for the segmentations of brain MR image with noise, we analysis the major drawbacks of conventional fuzzy clustering algorithm and the ideas of local information construction for current improved methods. Then we introduce a new method to compute the weights of the neighborhood which makes the pixels in the neighborhood have anisotropic weights. Mean-while, the algorithm is accelerated with histogram of the image. The experimental results show that our method has stronger anti-noise property and higher segmentation accuracy.(2) A modified possibilistic fuzzy c-means clustering algorithm is presented by combining the local and global intensity information. Based on the discussion of the strategies for the segmentations of brain MR image with intensity inhomogeneity, we focus on the segmentation based intensity inhomogeneity correction methods. To estimate the intensity inhomogeneities in the image, the proposed algorithm introduces the global intensity into the coherent local intensity clustering algorithm and takes the local and global intensity information into account. The segmentation target therefore is driven by two forces, one induced by the coherent local intensity and the other by the coherent global intensity, to ensure the smoothness of the derived optimal bias field and improve the accuracy of the segmentations. Comparisons with other approaches demonstrate the superior performance of the proposed algorithm.(3) A framework with modified fast fuzzy c-means clustering for brain MR images seg-mentation is proposed to overcome the intensity inhomogeneity, noise and partial volume effect simultaneously and improve image segmentation performance. A new automated method for centroids initialization is proposed to overcome the impact of intensity inhomogeneity and noise during initializing the centroids. An adaptive method to incorporate the local spatial continuity is proposed to overcome the noise effectively and prevent the edge from blurring. The frame-work is accelerated with histogram, and utilizes a set of basis function to estimate the bias field in the image. The weights of the regularization terms, which make the segmentations more accuracy, are all automatically computed to avoid the manually-tuned parameter and reduce the iteration steps of the algorithm. The proposed framework is fast and robust, thereby allowing for fully automatic applications.(4) To improve the uncertainty description of the dataset, we introduce the rough sets and interval type-2fuzzy sets into the fuzzy clustering model, and propose a generalized rough fuzzy c-means algorithm and an interval possibilistic fuzzy c-means clustering algorithm. In the first algorithm, a novel hybrid rough fuzzy c-means algorithm is proposed for brain MR image segmentation. Each cluster is characterized by three automatically determined rough-fuzzy regions, and accordingly the membership of each pixel is estimated with respect to the region it locates. The importance of each region is balanced by a weighting parameter, and the bias field in MR images is modeled by a linear combination of orthogonal polynomials. The weighting parameter estimation and bias field correction have been incorporated into the itera- tive clustering process. Experimental results demonstrate that the proposed algorithm is more robust to the initialization, noise, and bias field, and can produce more accurate and reliable segmentations. In the second algorithm, we utilize the interval type-2fuzzy set, and focus on the representation and management of uncertainty which is present in both fuzzy memberships and possibilistic typicalities of the patterns associated with the varying of fuzzifiers. Therefore, we extend a pattern set to interval type-2fuzzy sets using two fuzzifiers for membership and two fuzzifiers for possibilistic. Consequently, the proposed algorithm can simultaneously overcome the drawbacks and inherit the advantages of current interval type-2fuzzy set based algorithms. Experiments demonstrate the advantages of the method over state-of-the-art.(5) To overcome the low contrast in brain MR images, we introduce the second-order statistics (local variance) into the fuzzy clustering model, and propose two improved algorithms. By introducing the local Gaussian distribution fitting energy into the fuzzy clustering model, a fuzzy local Gaussian distribution fitting model is proposed to segment brain MR images. The means and standard deviation of local Gaussian distributions are updated iteratively and varying spatially, therefore the proposed model is more adaptive. Then, a new local scale computing method is introduced to estimate the local variances for local Gaussian distributions. Then the adaptive scale fuzzy local Gaussian distribution fitting model is proposed to improve the robust of the algorithm over the initializations. The experimental results show the advantages of the method over the state-of-the-art.