A SVM-based framework for autonomous volumetric medical image segmentation using hierarchical and coupled level sets S. Li * , T. Fevens, A. Krzyz ˙ak Computer Science Department, Concordia University, Montre ´al, Que ´bec, Canada Abstract. A volumetric segmentation framework combining level set and support vector machine (SVM) is proposed and implemented. Both hierarchical and coupled level set methods are combined to achieve a fast and robust segmentation framework. To accelerate SVM classification, an information reduction scheme is used. The framework is able to accelerate level set based segmentation while solving the initial curve problem. The experimental results are very promising. Although only Chan–Vese level set methods and Samson’s coupled level set methods have been implemented to test the framework; the framework can be generally extended to any hierarchical and coupled level sets methods. In addition, the work can also be extended to thin structure segmentation. D 2004 CARS and Elsevier B.V. All rights reserved. Keywords: Volumetric segmentation; Level set; Support vector machine; Coupled level set; Variational methods; Energy minimization 1. Introduction Volumetric medical image segmentation plays a critical role in computer-assisted radiology and surgery. It is more challenging compared to other imaging problems due to the large variability in shapes, complexity of medical structures, several kinds of artifacts, restrictive scanning methods and large amount of data to process. The application of the level set in medical image segmentation has become extremely popular because of its ability to capture the topology and geometry of shapes in medical imagery. There are some related works that have been presented mainly for two-dimensional (2D) medical image segmentation. Lorigo et al. [7] used codimension-two geodesic active contours to segment tubular structures. Deschamps [5] used region competition introduced in Ref. [16] for a similar purpose, while Fast Marching [10] was used in Ref. [12]. In Ref. [3] and Ref. [13], Chan and Vese proposed a method using the Mumford – Shah model for both 2D and volumetric segmentation. Later, a hierarchical scheme was used to extend this 0531-5131/ D 2004 CARS and Elsevier B.V. All rights reserved. doi:10.1016/j.ics.2004.03.349 * Corresponding author. Concordia University, Computer Science, 1400 De Maisonneuve Boulevard, West, H3G 1M8, Montre ´al, QC, Canada. Tel.: +1-514-3629300. E-mail address: shuo _ li@cs.concordia.ca (S. Li). www.ics-elsevier.com International Congress Series 1268 (2004) 207 – 212