Automatic 3D Liver Segmentation Using Sparse Representation of Global and Local Image Information via Level Set Formulation Saif Dawood Salman Al-Shaikhli 1 Michael Ying Yang 2 and Bodo Rosenhahn 1 1 tnt Institute for Information Processing / Leibniz University Hannover, Germany, 2 Computer Vision Lab. / TU Dresden, Germany Abstract. In this paper, a novel framework for automated liver seg- mentation via a level set formulation is presented. A sparse represen- tation of both global (region-based) and local (voxel-wise) image infor- mation is embedded in a level set formulation to innovate a new cost function. Two dictionaries are build: A region-based feature dictionary and a voxel-wise dictionary. These dictionaries are learned, using the K-SVD method, from a public database of liver segmentation challenge (MICCAI-SLiver07). The learned dictionaries provide prior knowledge to the level set formulation. For the quantitative evaluation, the pro- posed method is evaluated using the testing data of MICCAI-SLiver07 database. The results are evaluated using different metric scores com- puted by the challenge organizers. The experimental results demonstrate the superiority of the proposed framework by achieving the highest seg- mentation accuracy (79.6%) in comparison to the state-of-the-art meth- ods. 1 Introduction The liver is among the most common human organs to undergo invasive surgeries. In the case of liver tumors, an accurate 3D liver segmentation is important because a resection of the liver has to be carefully planned in order to preserve as much of the liver as possible [1]. Usually, computer tomography (CT) images are acquired for these purposes. The challenging aspects of liver segmentation in CT scan images can be summarized as follows: Firstly, the overlapping boundaries between the liver and surrounding organs such as heart, stomach, right kidney, and spleen, as illustrated in Fig. 1. Secondly, the large variability in liver shape, intensity distributions, and geometric properties from patient to patient, which make it difficult to describe the liver with model-based approaches [1]. Finally, liver segmentation, using a slice-by-slice approach, in 2D space is time consuming and gives inaccurate results. Therefore, for accurate segmentation results, volume segmentation methods in 3D space are more efficient [1]. In recent years, a variety of methods have been proposed to segment the liver. In 2007, liver segmentation competitions from CT data were held in conjunction with MICCAI [2]. Between 2007-2014, more than 35 automatic methods were arXiv:1508.01521v1 [cs.CV] 6 Aug 2015