Image local information is crucial for accurate segmentation of images with inhomogeneities in the intensity. In many popular methods, however, local regional information is normally underestimated and not included in the segmentation framework. In this paper, a segmentation is formalised as Bayesian Inference procedure. By considering the spatial variation of the intensity distribution, a likelihood that contains a joint distribution of intensity and spatial location is obtained. After incorporating this likelihood in a Bayesian maximum a posteriori estimation, we transformed the stochastic model into a segmentation method which utilises local information to segment images with inhomogeneities so as to guarantee a global optimum for the two-region case(foreground/background). Taking the computational complexity of our model into account, we take advantage of a GPU parallel algorithm to accelerate computation without losing segmentation accuracy. We demonstrate that taking into account local image information, our method results in significant improvements for image segmentation.