Image processing and Display


Purpose Segmentation of organs from a 3D medical volume is crucial in computer-assisted diagnosis and therapy. A number of segmentation algorithms based on binarisation, region growing, a deformable model, a level set and graph cut (GC) are available. GC can solve a variety of segmentation problems by minimising an energy function based on both a boundary and a region. Boykov et al. presented an algorithm for interactive segmentation that requires a few user interactions [1]. The salient feature of a GC-based algorithm is that it guarantees a global optimum of submodular functions. However, in some cases, a GC-based algorithm fails in segmentation due to noise and weak boundary contrast. To solve this problem, Freedman et al. introduced a shape prior in the GC [2]. Although they presented promising results, their method requires human interactions for specifying inside and outside voxels as well as estimating shapes. Moreover, their results sometimes included errors caused by strong noise and heavy blurring in a 3D medical volume. This paper presents an automated algorithm for shape prior based segmentation of organs in a 3D CT volume with GC that does not require any human interaction. It combines an automated shape estimation process and GC with a shape prior. An organ’s shape is estimated by a preliminary segmentation followed by shape estimation with a statistical shape model. In addition, this algorithm includes two novel shape energies. Its effectiveness is demonstrated by applying it to lung and liver segmentation in a 3D CT volume. Methods This paper formulates a segmentation problem as an energy minimisation problem in which the goal is to find a set of labels A = (A1, A2, ..., Ap ..., A|P|) that minimises an energy E(A) given by


6 Figures and Tables

Download Full PDF Version (Non-Commercial Use)