Automatic Seed Placement for Breast Lesion Segmentation on US Images Joan Massich 1,⋆ , Fabrice Meriaudeau 2 , Melcior Sent´ ıs 3 , Sergi Ganau 3 , Elsa P´ erez 4 , Robert Mart´ ı 1 , Arnau Oliver 1 , and Joan Mart´ ı 1 1 Computer Vision and Robotics Group, University of Girona, Spain jmassich@atc.udg.edu 2 Laboratoire Le2i-UMR CNRS, University of Burgundy, Le Creusot, France 3 Department of Breast and Gynecological Radiology, UDIAT-Diagnostic Center, Parc Taul´ ı Corporation, Sabadell, Spain 4 Department of Radiology, Hospital Josep Trueta of Girona, Spain Abstract. Breast lesion boundaries have been mostly extracted by us- ing conventional approaches as a previous step in the development of computer-aided diagnosis systems. Among these, region growing is a frequently used segmentation method. To make the segmentation com- pletely automatic, most of the region growing methods incorporate auto- matic selection of the seed points. This paper proposes a new automatic seed placement algorithm for breast lesion segmentation on ultrasound images by means of assigning the probability of belonging to a lesion for every pixel depending on intensity, texture and geometrical constraints. The proposal has been evaluated using a set of sonographic breast im- ages with accompanying expert-provided ground truth, and successfully compared to other existing algorithms. Keywords: seed placement, ultrasound, segmentation, breast cancer. 1 Introduction Breast cancer constitutes a leading cause of death for women in developed coun- tries, and is most effectively treated when diagnosed at an early stage[8]. Digital Mammography is currently the most powerful screening tool for breast cancer [5], although ultrasound images can provide useful complementary information in cases where a tumor presence can be shielded due to dense glandular breast tissue [9]. Despite ultrasound imaging is a non-expensive and non-invasive tech- nique with no side effects, its use in CAD systems is still under development. A feasible explanation is that performing automatic segmentation in US images is currently a challenge because they often suffer from poor quality and tend to generate artifacts: weak edges due to acoustic similarity between adjacent tissues, shadows as a consequence of the signal attenuation preventing to screen ⋆ This work was partially supported by the Spanish Science and Innovation grant nb. TIN2011-23704, the Regional Council of Burgundy and the University of Girona BR grant nb. 09/22. A.D.A. Maidment, P.R. Bakic, and S. Gavenonis (Eds.): IWDM 2012, LNCS 7361, pp. 308–315, 2012. c  Springer-Verlag Berlin Heidelberg 2012