Mammographic Masses Segmentation Using Implicit Deformable Models: The LCV Model in Comparison with the Osher-Sethian Model Fouzia Boutaouche and Nacera Benamrane Department of Informatics, University of Science and Technology of Oran Mohamed BOUDIAF, USTO-MB, Oran, Algeria Email: boutaouche-f@netcourrier.com, nacera-benamrane@univ-usto.dz Abstract—Breast cancer is one of the leading causes of cancer death among women. As such, the role of digital mammographic screening is to detect cancerous lesions, at an early stage, and to provide high accuracy in the analysis of the size, shape, and location of abnormalities. Segmentation is arguably one of the most important aspects of a computer aided detection system, particularly for masses. This paper attempt to introduce two level set segmentation models for mass detection on digitized mammograms. The first in an edge-based level set algorithm, proposed by Osher and Sethian. The second is a region- based level set algorithm called the local Shan-Vese model. A comparative study will be given, in which we will assess the performance of the second approach in terms of efficiency. Index Terms—breast segmentation, Level Set method, local chan-vese model, Osher and Sethian algorithm I. INTRODUCTION A mammogram is an x-ray picture of the breast. It’s considered as the most common method for early detection of breast cancer. The earliest sign of breast cancer is an abnormality detected on a mammogram. It can appear as an abnormal area of density mass, or calcification. Masses are space-occupying lesions, described by their shapes, margins, and denseness properties. Interpretation of mammograms can be difficult, because some breast cancers are hard to visualize, this is due to the nature of the beast, the location and the size of the abnormality. Furthermore, masses can have unclear borders, and can be obscured by glandular issues [1]. Accuracy of segmentation is crucial because many features extracted from segmented regions are used to discriminate benign and malignant lesions. Classical approaches to solve segmentation are divided in different categories: histogram thresholding [2], region-based methods [3], model-based methods (active contour, level set, Markov random field) [4], [5], clustering methods [6], [7]. In this paper, we introduce a robust and an efficient segmentation method from the geometric models family, for performing contour evolution to extract masses: the Manuscript received May 21, 2014; revised November 28, 2014. level set method, also called the implicit deformable model. The fundamental idea is to evolve the contour in such a way that it stops on the boundaries of the foreground region [8]. To perform the contour evolution, two types of forces are computed: the internal forces defined to keep the model smooth during the contour evolution process, and the external forces defined to move the contour toward the boundary of an object [9], [10]. In level set algorithms, we are interested to segment a single part from the whole image; this kind of methods is called image selective segmentation. Segmenting a single region aims to isolate a suspected abnormality in a mammogram, in order to extract relevant features (surface, perimeter, texture, variance, entropy…) used to classify breast masses. Among the level set models, we will focus on two models: the Osher and Sethian model [9] which is an edge-based level set model, and the local Chan-Vese model which is regarded as a region-based level set model. In Osher and Sethian model, the curve evolution is guided by the gradient of the image to stop the evolving curve on the boundary of the desired object. We will see that this solution suffers from several problems in detecting masses and curves may pass through the true boundaries. The Local Chan-Vese model [11] incorporates region-based information into the energy functional to stabilize the evolution to local variations. The functional energy is based on three terms: the global term, which includes global properties as the intensity average, the local term which incorporates local statistical information to improve the segmentation process, and the regularization term, used to ensure curve evolution stability. Experimental results show that this method is an efficient and accurate method to isolate and extract masses in mammograms, with weak boundaries, and intensity inhomogeneity, especially, when the abnormality presents physical characteristics similar to those of normal tissue, with blurred contour or hidden by superimposed or adjacent normal tissue, with an inhomogeneous density. This is clearly visible when the breast is dense. Dense breasts can make mammograms harder to interpret because both tumors and dense breast tissue appear white. The segmentation model we propose 100 ©2014 Engineering and Technology Publishing doi: 10.12720/joig.2.2.100-105 Journal of Image and Graphics, Volume 2, No.2, December 2014