doi:10.1016/j.ultrasmedbio.2005.07.005 ● Original Contribution ULTRASOUND IMAGE SEGMENTATION USING SPECTRAL CLUSTERING NECULAI ARCHIP,* ROBERT ROHLING, † PETER COOPERBERG ‡ and HAMID TAHMASEBPOUR ‡ *Harvard Medical School, Brigham and Women’s Hospital, Boston, MA, USA; † Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, B.C., Canada; and ‡ Department of Radiology, University of British Columbia, Vancouver, B.C., Canada (Received 14 Feburary 2005; in final form 7 July 2005) Abstract—Segmentation of ultrasound images is necessary in a variety of clinical applications, but the develop- ment of automatic techniques is still an open problem. Spectral clustering techniques have recently become popular for data and image analysis. In particular, image segmentation has been proposed via the normalized cut (NCut) criterion. This article describes an initial investigation to determine the suitability of such segmentation techniques for ultrasound images. The adaptation of the NCut technique to ultrasound is described first. Segmentation is then performed on simulated ultrasound images. Tests are also performed on abdominal and fetal images with the segmentation results compared to manual segmentation. The success of the segmentation on these test cases warrants further research into NCut-based segmentation of ultrasound images. E-mail: (narchip@bwh.harvard.edu). © 2005 World Federation for Ultrasound in Medicine & Biology. Key Words: Image segmentation, Spectral clustering, Fiedler eigenvector, Graph partitioning. INTRODUCTION The segmentation of medical images has been a chal- lenging research topic for many years. Successful seg- mentation techniques can be used in a variety of clinical applications. Some example applications are the mea- surement of the volume of cancer tumors and cysts (Pathak et al., 2000), the planning of radiotherapy pro- cedures (Archip et al., 2002), and the reconstruction of three-dimensional anatomical models of patients for minimally invasive surgery (Kaus et al., 2001). In clin- ical practice, ultrasound images are often segmented manually, but manual techniques can be laborious. More sophisticated techniques are needed. Semiautomatic segmentation techniques are a first step towards providing computer assistance, whereby the user initializes or guides the segmentation process. A common semiautomatic technique is based on active contours, such as “snakes” in 2D or “active surfaces” in 3D (Kass et al., 1988). Active contours are usually ini- tialized near an organ boundary, and the process auto- matically deforms the contour toward the boundary. For example, an active contour model that uses several tem- porally adjacent images during the extraction of the tongue surface from a sequence of images is presented by Akgul et al. (2000). Another active contour approach (Hamarneh and Gustavsson, 2000) extracts the left ven- tricle of the heart in echocardiography by including a priori knowledge of ventricle shape to restrict the allow- able range of deformations. Another approach uses model-based initialization and a discrete dynamic con- tour to segment the prostate (Ladak et al., 2000). The algorithm requires the user to select four points on the organ boundary from which an estimate of the prostate shape is interpolated by using cubic functions. The esti- mated contour is then used as an initialization stage for an active contour to better fit the prostate boundary. An extension of this algorithm, which uses six user selected control points to initialize a 3D active surface, is pro- posed by Hu et al. (2002). The active surface is then used to segment 3D ultrasound volumes of the prostate. A 3D deformable surface is also used by Ghanei and Soltanian- Zadeh (2002) for segmenting the prostate in 3D ultra- sound. The user must first outline the prostate in several 2D cross-sections of the 3D volume data to initialize the model. Jacob et al. (2001) use snakes with a temporal Kalman filter to extract periodic motions from a se- quence of ultrasound images. The algorithm performs an Address correspondence to: Dr. Neculai Archip, Harvard Medi- cal School, Brigham and Women’s Hospital, Thorn 329, Dept Radiol- ogy, 75 Francis St, Boston, MA, 02115, E-mail: narchip@bwh. harvard.edu Ultrasound in Med. & Biol., Vol. 31, No. 11, pp. 1485–1497, 2005 Copyright © 2005 World Federation for Ultrasound in Medicine & Biology Printed in the USA. All rights reserved 0301-5629/05/$–see front matter 1485