Graph Cuts Segmentation with Statistical Shape Priors For Medical Images Jie Zhu-Jacquot School of Electrical and Computer Engineering Cornell University Ithaca, NY 14853 Email: jz85@cornell.edu Ramin Zabih Department of Computer Science Department of Radiology Cornell University Ithaca, NY 14853 Email: rdz@cs.cornell.edu Abstract Segmentation of medical images is an important step in many clinical and diagnostic imaging applications. Medi- cal images present many challenges for automated segmen- tation including poor contrast at tissue boundaries. Tra- ditional segmentation methods based solely on information from the image do not work well in such cases. Statistical shape information for objects in medical images are easy to obtain. In this paper, we propose a graph cuts-based seg- mentation method for medical images that incorporates sta- tistical shape priors to increase robustness. Our proposed method is able to deal with complex shapes and shape vari- ations while taking advantage of the globally efficient opti- mization by graph cuts. We demonstrate the effectiveness of our method on kidney images without strong boundaries. 1 Introduction Medical image segmentation, i.e. assigning voxels in medical images to tissues classes, is a first step to vol- ume analysis and has many research and clinical applica- tions. These applications usually involve a vast amount of data, therefore segmentation by human experts can be time- consuming. Data segmented by human experts also tends to show inter- and intra-observer inconsistency [1]. For these reasons, automated segmentation of medical images is of great importance and interest. Medical images present many challenges for automated segmentation. For example, tissue boundaries in medical images sometimes have poor contrast due to the overlap in intensity ranges for multiple tissues. Segmentation meth- ods based on information from the image alone do not work well in such cases and additional constraints are necessary. Statistical shape information for objects in medical im- ages is easy to obtain because the objects do not vary dras- tically from scan to scan and are generally imaged from the same perspectives. Level set-based segmentation meth- ods using statistical shape information have been studied [2, 3, 4, 5]. The graph cuts-based segmentation method has recently become popular because it allows for a globally optimal ef- ficient solution in an N-dimensional setting [6]. Despite its advantages, graph cuts cannot produce an accurate segmen- tation for images with weak boundaries. There have been recent attempts to add a shape prior to the graph cuts seg- mentation technique. [7] proposed the usage of an elliptical prior, which provides a fast solution but can only be applied when the desired object resembles an ellipse. [8] presented a method that uses a fixed shape template aligned with the image by the user input. The effectiveness of this method is limited to when the fixed template and the desired segmen- tation somewhat match. In this paper, we propose a novel segmentation method that incorporates statistical shape priors to the graph cuts technique for robust and accurate segmentations of med- ical images. We introduce novel terms accounting for global/local shape properties to the graph cuts representa- tion. Our proposed method is able to deal with complex shapes and shape variations while taking advantage of the globally efficient optimization by graph cuts. 2 Segmentation Method In this section, we first give a brief summary of the graph cuts image segmentation framework. We then discuss our models for shape representation and intensity probability distribution. We end by presenting our novel objective func- tion and its optimization. 2.1 Graph Cuts The basic graph cuts image segmentation framework is developed in [9]. The idea is as follows. An image is Copyright © SITIS - 588 -