Maximum a posterior based level set approach for image segmentation with intensity inhomogeneity

Abstract

Intensity inhomogeneity is an unavoidable obstacle in image segmentation, which causes inaccuracy in object extraction. Generally, the approach to tackling intensity inhomogeneity is constructing a bias field descriptor which may lead to corruption of image intensity. In this paper, we propose a novel level set model based on maximum a posterior principle. To properly collect the objectslocal information, the proposed method utilizes Gaussian distribution to model the conditional probability of image intensity within specific patches. To construct the prior information, we then model the intensity inhomogeneity as Gaussian distribution whose mean is 1 and whose variance is the same as image intensity. Finally, the maximum a posterior based energy functional combined local image information and adequate prior information is defined. In addition, our method can be adopted and transformed into the state-of-the-art methods. To validate its effectiveness and performance, we compare our method with popular deep learning methods and classical level set methods. The roubstness analysis of initial contour, noise and intensity bias is given. Experimental results show our method achieves outstanding adaptability and stability.