A robust active contour initialization and gradient vector flow for ultrasound image segmentation C Tauber, H Batatia, A Ayache IRIT-ENSEEIHT, Toulouse, France Abstract Speckle and low contrast make ultrasound image segmentation a difficult task. This paper presents an original robust active contour energy and the corresponding quasi-automatic initialization. Both are based on the coefficient of variation gradient vector flow. Our approach combines anisotropic diffusion with the gradient vector flow field. The gradient vector flow is calculated from a map of the amplitudes of the coefficient of variation. This makes it more robust to speckle. The centers of divergence are calculated and used to initialize the active contour model. The method has been tested on different echocardographic images. The results presented are very encouraging. 1. Introduction Ultrasound (US) imagery is characterized by low signal to noise ratio, low contrast between tissues and speckle contamination causing erroneous detection of cavities boundaries. Active contour models (snakes) deal with some of these limitations. They consider boundaries as inherently connected smooth curves. A snake is a curve that evolves from an initial position towards the boundary of an object, minimizing some energy functional [1, 2, 3]. A B-spline snake is an energy minimizing spline parameterized by its control points. The smoothness of the snake is implicitly given by the B-spline model [4, 5, 6]. The energy consists in two terms: the internal energy and the external energy. The first affects the smoothness of the curve, and the second attracts the snake toward image features. A number of external energy terms have been proposed [7, 8, 9] . Most of these approaches either use gradient information or global image statistics. Among them the gradient vector flow field introduced by Xu [9] has the inherent property of being able to reconstruct subjective contours. These contours are edges that are not actually present in an image, but are perceived nevertheless. This characteristic is very attractive for US imagery where connected boundaries are less likely to be found. However this method cannot be used efficiently for US imaging because of the presence of speckle. A novel method for anisotropic diffusion of ultrasound images was introduced in [10]. It uses the local coefficient of variation (LCV) [11] and a robust diffusion tensor to filter echographic images. The LCV is computed locally and compared to the global coefficient of variation (GCV). In homogeneous areas affected by speckle LCV is close to GCV. Near edges LCV becomes greater. This diffusion technique was successfully used in [12] as a pre-step for segmentation of the heart cavities in US images. However this B-spline snake method is difficult to initialize. It is well known that the snake initialization accuracy influences significantly the segmentation. Using primer contour as the snake initialization have been proposed in [13, 14], multi-scale approach was proposed in [15], balloon snakes and GVF were proposed in [2, 9]. In this paper we derive a new gradient vector flow based on the amplitude of the coefficient of variation, called Speckle resistant Gradient Vector Flow (s-GVF). First we use a robust anisotropic model to filter the US image. Then we generate an image of the LCV amplitude which is used to generate the s-GVF. The s-GVF centers of divergence are used to develop a quasi-automatic curve initialization and the s-GVF is used to attract the B-spline snake toward the cavities boundaries. The remainder of this paper is organized as follows. We present the s-GVF in section 2.1. In section 2.2 we present a new model for quasi-automatic initialization of the B- spline snake. The results are shown in section 3, and a discussion and conclusion can be found in section 4. 2. Methods 2.1. The speckle resistant gradient vector flow To generate the s-GVF we first apply anisotropic diffusion described in [10] to generate a coefficient of variation amplitude map. The anisotropic diffusion is MVA2005 IAPR Conference on Machine VIsion Applications, May 16-18, 2005 Tsukuba Science City, Japan 4-1 164