Research On Algorithms Based On Fuzzy And Random Models For Magnetic Resonance Brain Image Segmentation

There are two purposes for the segmentation of MR brain images. The first one is to segment MR brain images into different tissue classes, especially gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF), which is crucial to the image registration, 3D reconstruction and medical image visualization. The second one is to extract the focal region of interesting (ROI) from other tissues in order to assist physicians in making right diagnosis, and working out the therapeutic strategy.The research on MR brain image segmentation has been an important field in medical image processing and analysis. There are a number of factors that cause current segmentation algorithms fail to satisfy the need of clinical practice, including 1) the individual differences in the tissue anatomy; 2) slow calculating speed and inaccuracy; and 3) poor image quality affected by noise, intensive inhomogeneity and partial volume effect (PVE), etc.Medical images behave fuzziness duo to PVE artifacts and the uncertainty in some focal regions. The idea of using membership function associated with fuzzy-set theory to represent partial volume proportions of each "pure" tissue has been a quite popular and widely used model, in which Fuzzy c-means (FCM) clustering algorithm is the well-established approach to the implementation of the image segmentation. However, the conventional FCM fails to incorporate the spatial information of the image leading to aberrant consequences in the case of dealing with low signal-to-noise ratio (SNR) MR images.The Markov random field (MRF) segmentation has been successfully applied to this issue in the presence of noise by taking into account a priori knowledge of the spatial correlations of the image using Gibbs distribution and maximum a posteriori (MAP). However, there are still problems associated with MRF, such as the difficulty in dealing with fuzzy characteristics of images, parameter estimation, and the tendency of over-segmentation.In this paper, we look into the segmentation algorithms based on fuzzy set and MRF, and the main contributions to which are as follows: 1) a modified FCM clustering algorithm is proposed to improve the segmentation accuracy of FCM; 2) a Fuzzy Markov random field (FMRF) model is introduced by combining fuzzy set and MRF, and the modeling, parameter estimate, optimization methods and algorithm about FMRF are studied; 3) a fuzzy connectedness-based segmentation method of multiple sclerosis (MS) lesions and a MRF-based segmentation method of MS lesions are developed; 4) inhomogeneous MRF (IMRF) model is studied and the parameter in IMRF is estimated using fuzzy connectedness. The performance of these algorithms is remarkably superior to the conventional ones in terms of accuracy and robustness.In chapterâ…¡, we present a modified FCM clustering algorithm for brain image segmentation with a membership smoothing constraint (MC-FCM). The rationale of which is that in general an ideal MR image is assumed to be piecewise constant, a membership smoothing constraint is therefore appended to the object function of conventional FCM so as to incorporate the spatial information of the image. The new mathematical programming formula can thus be solved by the Lagrange multiplier. The validity of this algorithm is evaluated using simulated brain MR images with different noise level and real brain MR image. The results show that MC-FCM is overperformed than the conventional counterparts, and is as well as simple, fast, and robust.In chapterâ…¢, we inspect the properties of both fuzzy set and MRF, then introduce the notion of fuzzy membership to the conventional MRF model---therefore forms a Fuzzy Markov random field (FMRF) model. Applying MAP method to the segmentation problem leads to a mathematical programming problem, which can be solved by deriving the formula of determining the membership values for each voxel to indicate the partial volume degree. The results obtained by testing both simulated and clinical data, show that FMRF can segment them more accurately than the conventional model-based and FCM do as well.In chapterâ…£andâ…¤, we investigate the segmentation of MS lesions --- an inflammatory demyelinating disease that would damage central nervous system. There is a growing attention to this area for the conventional segmentation algorithms are not working well due to the effects of noises, intensive inhomogeneities, the behavior of MS lesions etc.In chapterâ…£, An automatic segmentation algorithm of MS lesions for MR FLAIR images is presented based on fuzzy connectedness by using a priori knowledge of characteristics of MS lesions and anatomical structures. The connectedness employs fuzzy relationship to describe the analogy of two neighboring pixels. The segmentation of focal regions is accomplished by choosing pixel-seeds, and then expanding regions using connectedness between the pixels. The pixel-seed selection is done automatically to facilitate practical applications. A novel brain tissue extraction algorithm is also presented using region expanding method as the preprocessing, which is able to automatically remove tissues other than those of brain, such as skull, scalp etc. The testing results using clinical MR FLAIR brain images demonstrate that the performance of the proposed algorithm is significantly improved over the conventional FCM clustering and MRF model-based algorithms. This unsupervised algorithm can be used in clinical practice with adequate calculating speed, and robustness.In chapterâ…¤, we develop a MRF-based algorithm for MS lesions segmentation by utilizing the morphological characteristics of MS lesion tissues. The regions circumscribed by white matter are extracted at first by MRF segmentation and region growing methods; the abstracted regions are then segmented again using MRF algorithm. The testing results for T₂-weighted MR brain images show the proposed algorithm is robust and accurate enough for clinical use.In chapterâ…¥, A improved unsupervised algorithm for image segmentation is proposed using an inhomogeneous MRF model, in which the parameter is estimated in fuzzy connectedness. The conventional MRF-based segmentation algorithms always assume the MRF is homogeneous, i.e., the pixels are uncorrelated spatially. This assumption is usually not satisfied in practice, which requires a more precise model like inhomogeneous MRF to represent the context information. Simulated brain MR images with different noise level and real brain MR images are tested. The results show that the proposed algorithm is more accurate in comparison with others.