Extracting the pectoral muscle in screening mammograms using a graph pyramid Fei Ma Mariusz Bajger Murk J. Bottema School of Informatics and Engineering, Flinders University PO Box 2100, Adelaide SA, 5001 email: ma0029@infoeng.flinders.edu.au Abstract A graph based method is introduced to segment the pec- toral muscles in screening mammograms. An adaptive pyramid is used to segment the mammogram into a num- ber of components. Components forming the pectoral mus- cle are identified based on position, intensity, and shape. The boundary of the union of these components forms an initial boundary that is refined via an adaptive deformable contour model. The method is tested on 83 medio-lateral oblique mammograms from the Mini-MIAS database. Seg- mentation results are evaluated in terms of the proportion of correctly assigned pixels. Performance compares well with existing methods based on Hough transform and on Gabor wavelets. 1 Introduction Breast cancer screening programs based on mammogra- phy are used in many countries to facilitate early detection of breast cancer. Normally mammograms are evaluated vi- sually by radiologists for signs of cancer. Since the mid 1980’s, many computer algorithms have been proposed for automating various aspects of detecting the presence of can- cer in mammograms and commercial products now exist that implement some of these programs. While detection rates for automatic systems are quite high, the false pos- itive detection rates are also high. Accordingly, work con- tinues on improving all aspects of computer-aided detection (CAD) for mammography. Accurate segmentation of the pectoral muscle is among the many tasks that is needed to improve CAD for mam- mography. The pectoral muscle is one of the few anatomical features that appears clearly and reliably in medio-lateral oblique (MLO) view mammograms. The pectoral muscle is an important landmark both for providing contextual infor- mation regarding anatomies and for image registration. To a first approximation, the pectoral muscle appears as a bright triangular patch in the upper left or upper right cor- ner (depending on right or left breast) of the image. This motivated initial algorithms based on the Hough transform [2] [4]. The pectoral muscle is usually not exactly triangu- lar and more accurate segmentation was achieved by using Gabor wavelets to segment the pectoral muscle without as- suming straight boundaries [5]. Aside from incorporating general shape and location assumptions of pectoral muscle, these methods rely only on local image information. In this paper, graph theory methods are used in an ef- fort to incorporate global image information in segmenta- tion. Graph pyramids were introduced by Tanimato and Pavlidis in 1975 [6] and have been applied widely in im- age processing. A graph pyramid is a stack of successively reduced graphs. At each level in the stack, the graph is a reduction of the graph at the previous level. A vertex of a graph at one level is connected to a number of vertices at the previous level. The vertex in the higher level is called the parent of the vertices in the previous level and the set of vertices to which the parent is connected in the previ- ous level (the children) is called the receptive field of the vertex. The collection of graphs forms a multi-resolution description of the image, but unlike multi-resolution rep- resentations via wavelets or filter banks, the connectivity between layers provides a vehicle for tracking information from disparate regions of the image. The connectivity be- tween layers may be based directly on image intensities or derived image properties, thus providing a flexible tool for associating information content. This paper is arranged as follows. In section 2, adap- tive pyramids (AP) are described in detail. In section 3, a method for extracting the pectoral muscle, including an adaptive deformable contour model to refine the pectoral muscle boundary, is presented and in section 4 the perfor- mance of the method on a standard set of mammograms is reported.