New Approaches To Segmentation Of Brain MR Images Based On Gibbs Random Field Theory

Image segmentation, partitioning an image into different regions with some specific properties and labeling each pixel with its underlying class, has always been an important and challenging problem for many years. The main difficulties lie in the great variability of images and the presence of noises. It is widely accepted that the key for a better segmentation is to incorporate high-level a-priori knowledge into the framework of segmentation. In this thesis, traditional segmentation algorithms are discussed and a-priori knowledge is incorporated through the Gibbs random field (GRF) theory to deal with a particularly important problem in the area of medical image analysis: automatic segmentation of brain MR images.Generally, the finite mixture (FM) model is the most widely used model for statistical segmentation of brain MR images because of its effective modeling of intensity distribution and its simple mathematical form. However, being a kind of histogram-based model for segmentation, the FM model has an intrinsic limitation-wo spatial information is taken into account. This causes the FM model to work only on well-defined images with low levels of noise; unfortunately, this is not often the case in reality. In this thesis, we present an improved segmentation algorithm: A-priori spatial knowledge is incorporated into the FM model through the GRF theory. Furthermore, the new algorithm is also an automatic one with the model parameters estimated by EM algorithm and initialized by the tree-structure AT-means algorithm.The experiments on simulated MR images and real medical MR images prove that the proposed algorithm is insensitive to noises and can more precisely segment brain MR images into three different tissues: gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF).Finally, Fuzzy C-means (FCM), one of well-known unsupervised clustering methods, is discussed. Similar to the FM model, the standard FCM algorithm, when used for image segmentation, takes no account of spatial information either. In this thesis, refusable level is defined based on GRF theory and incorporated into traditional FCM. In our experiments, it is shown to be effective for this new GFCM algorithm to segment images degenerated by noises.