Automatic Multiple Active Contour Initialization Using Multi-stage Evolution Khwunta Kirimasthong School of Information, Computer and Communication Technology, Sirindhorn International Institute of Technology, Thammasart University Pathum Thani, Thailand khwunta.kir@mfu.ac.th Annupan Rodtook Department of Computer Science Ramkhamhaeng University Bangkok, Thailand Abstract— The multiple active contours (snakes) are widely used in image processing to extract the feature of interest. However, initialization of the active contours is still an open problem. This paper proposes a new completely automatic initialization of snakes based on a combination of Phrase Portrait Analysis (PPA), K-means clustering and multi-stage evolution of the active contours. The proposed method has been tested with both synthetic and real ultrasound images of breast. The experiments show that the proposed automatic initialization is applicable to noisy ultrasound images and works well to extraction the tumor boundary. Keywords— Automatic initialization; Medical image processing; Multiple active contours; Phrase portrait analysis I. INTRODUCTION The computer aided diagnostic system for segmentation medical images can help the physicians to pre-locate the features of interest and save time. The accuracy of the computer based diagnostics of ultrasound (US) images is still not sufficient [1]. It is often difficult to separate the tumor region from the background even when the existence of the tumor is evident. Among the most promising techniques for segmentation of complex objects from digital images are active contours or snakes, originally introduced by Kass et al. [2]. The active contour tries to minimize the functional energy. The classical snake performance improved to higher-order parametric model by Rochery et al. [3]. The multiple quadratic active contours are purposed to extract the complex road from satellite images [4]. However, the success of the snake-based segmentation depends strongly on initialization that is on the position where the snake is initially placed. The existing works initial the snake close to the actual boundary and manually initialization [4-5]. Moreover, the snake segmentation is enhanced by improving the external force field. Xu and Prince proposed the gradient vector flow (GVF) [6] that increases the capture range of gradients far from boundary, and the generalized gradient vector flow (GGVF) that improves GVF by introducing the weighting function to replace the diffusion coefficient [7]. Furthermore, PPA is applied in preprocessing step to improve the vector flow field [5]. PPA detects linear flow configurations and classifies them as the object boundary, noise or the background (regular point) [5]. Up to now several solutions of the active contour initialization problem have been proposed and implemented. A quasi-automatic initialization that relies on the divergence of gradient vector flow field is proposed in [8]. The mean shape algorithm for automatic initialization is used in [9]. The Poisson inverse gradient for automatic active contour initialization is introduced in [10]. However, the proposed methods do not show proper results when applied to US images of the breast cancer. Therefore, the objective of this paper is to purpose a completely automatic initialization model based on the combination of PPA, clustering algorithm, and multiple active contours applied using a multi-stage evolution algorithm for extracting the breast tumor cancer from US image. Due to the US image is so complex and noisy or sometime has the other small issues. Therefore, to avoid the active contour segmentation getting stuck into the local region of noise or other small issue, this work will automatic initialize the multiple snakes around the convergence or the regions of noise by applying the clustering algorithm that is K-mean clustering to group the seed points for multiple initial snakes. II. METHODLOGY In this work, the preprocessing sequence includes a speckle noise reduction, median filtering, contrast enhancing and region growing. The PPA [5] is combined with edge-based detection in the preprocessing sequence to improve the general gradient vector flow (GGVF) as the resulting image (Fig. 2(d)). The PPA detects the linear flow pattern of boundary (Fig. 2(b)). The linear flow pattern of noise is detected and smoothed (Fig. 2(c)). The grey-level image intensities are mapped while PPA preprocessing shown as Fig. 3(a). Then, the initial multiple active contours based on the combination of PPA, clustering algorithm, and multiple active contours will be generated automatically. Our proposed automatic initialization model for multiple snakes consists of six steps shown as Fig. 1. We evaluate the PPA map that is classified as the background, boundary and the noise. Next, we perform a geometric analysis of the pixels with a strong PPA grey-level map and set up seeds for multiple