IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 3, NO. 1, FEBRUARY 2009 135 Segmentation of Head and Neck Lymph Node Regions for Radiotherapy Planning Using Active Contour-Based Atlas Registration Subrahmanyam Gorthi, Valérie Duay, Nawal Houhou, Meritxell Bach Cuadra, Ulrike Schick, Minerva Becker, Abdelkarim S. Allal, and Jean-Philippe Thiran, Senior Member, IEEE Abstract—In this paper, we present the segmentation of the head and neck lymph node regions using a new active contour-based atlas registration model. We propose to segment the lymph node regions without directly including them in the atlas registration process; instead, they are segmented using the dense deformation field computed from the registration of the atlas structures with distinct boundaries. This approach results in robust and accurate segmentation of the lymph node regions even in the presence of significant anatomical variations between the atlas-image and the patient’s image to be segmented. We also present a quantitative evaluation of lymph node regions segmentation using various sta- tistical as well as geometrical metrics: sensitivity, specificity, dice similarity coefficient and Hausdorff distance. A comparison of the proposed method with two other state of the art methods is pre- sented. The robustness of the proposed method to the atlas selec- tion, in segmenting the lymph node regions, is also evaluated. Index Terms—Atlas-based segmentation, head and neck, IMRT, lymph node regions, non-rigid registration, radiotherapy. I. INTRODUCTION I NTENSITY-modulated radiotherapy (IMRT) is the ultimate high precision technique to accurately deliver X-ray radia- tion treatment for different tumor locations of the patients. How- ever, one of the significant obstacles in the widespread imple- mentation of IMRT, for head and neck (H&N) cancer, concerns the complexity of target definition. In the case of H&N carci- nomas radiotherapy, besides the gross tumor volume, the radi- ation oncologist has to segment the clinical target volume and the complicated planning target volume which contains different lymph node levels. Each lymph node level or group of levels Manuscript received April 15, 2008; revised October 15, 2008. Current version published February 19, 2009. This work was supported in part by the Swiss National Research Funds under Grant 3252B0-107873 and by the Center for Biomedical Imaging (CIBM) of the Geneva-Lausanne Universities and the EPFL, as well as the foundations Leenaards and Louis-Jeantet. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Jianhua Yao. S. Gorthi, N. Houhou, M. Bach Cuadra, and J.-P. Thiran are with the Signal Processing Laboratory (LTS5), Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland (e-mail: subrahmanyam.gorthi@epfl.ch; nawal.houhou@epfl.ch; meritxell.bach@epfl.ch; jp.thiran@epfl.ch). V. Duay is with the Institute of Bio-Engineering, University of Applied Sci- ences Western Switzerland (HES-SO Genéve), Geneva, Switzerland (e-mail: valerie.duay@hesge.ch). U. Schick, M. Becker, and A. S. Allal are with the Department of Radi- ation Oncology, University Hospital Geneva, Switzerland (e-mail: Ulrike. Schick@hcuge.ch; Minerva.Becker@hcuge.ch; Abdelkarim.Allal@hcuge.ch). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JSTSP.2008.2011104 correspond to a potential area of spread for a given tumor sub- location. Since the IMRT approach prerequisites the segmenta- tion of all the volumes to be treated as well as the organs at risk, it is easy to understand that its routine use for H&N tumors is not common at the present time. Besides the precise contouring of primary H&N tumors that is often difficult, the accurate, re- producible and time-efficient contouring of elective nodal risk regions represents an even greater challenge. The tentative to im- plement lymph nodes levels segmentation in the clinical environ- ment were initially based on the translation of the surgical lymph nodes levels to CT-based regions which meant meticulous seg- mentation of each CT regions on each slice of the planning CT scan, a laborious process that was considered as incompatible with a routine clinical practice [1]. Indeed, experienced H&N cancer specialists generally spend several hours to fully contour and refine desired targets for a single H&N IMRT case. In a study reported by Song et al. [2], the average physician working time to design a H&N treatment contours for the target definition was 2.7 h for IMRT approach compared to 0.3 h for the conventional three field plan. In summary, the major challenge in the routine clinical implementation of IMRT for H&N region is to delin- eate the lymph nodes automatically and accurately. Grégoire et al. [3] presented guidelines for delineating the lymph nodes in the H&N region. Fig. 4 shows the manually delineated lymph node levels: IA, IB-left, IB-right, IIA-left, IIA-right, IIB-left, IIB-right, III-left, III-right, IV-left, IV-right, VA-left, VA-right, VB-left, VB-right, and VI in the computed tomography (CT) images. 3-D volumes of these lymph nodes are shown in Fig. 8. Most of these lymph nodes do not have distinct boundaries; rather they are defined with respect to other distinct landmark structures in the H&N region and hence posing challenges in the automated segmentation. The lymph node segmentation techniques that have been re- ported so far can be broadly classified into two categories. The first category of techniques assume that at least a portion of the lymph node to be segmented has a distinct boundary with the surrounding structures. The second category of techniques do not assume the existence of any such distinct boundaries. In the first category, Rogowska et al. [4] used various basic techniques like threshold selection, sobel/watershed technique and deformable contour algorithm for the segmentation of lymph nodes. Their evaluation was on synthetic images. In [5], Honea et al. semi-automatically segmented the lymph nodes with slice-wise active contours and active surface models. Their eval- uation was also on synthetic images. In [6], Yan et al. proposed an 1932-4553/$25.00 © 2009 IEEE Authorized licensed use limited to: EPFL LAUSANNE. Downloaded on April 28,2010 at 11:25:19 UTC from IEEE Xplore. Restrictions apply.