GRANULOMETRIC FEATURE EXTRACTION FOR CANCEROUS TISSUE CLASSIFICATION IN BREAST IMAGES Lucia Ballerini Dept. of Technology ¨ Orebro University 70182 ¨ Orebro, Sweden lucia@aass.oru.se Lennart Franz´ en Dept. of Pathology ¨ Orebro University Hospital 70185 ¨ Orebro, Sweden lennart.franzen@orebroll.se ABSTRACT This paper presents an application to a medical problem of methods of image analysis based on mathematical mor- phology. The medical problem consists the detection of abnormalities in microscopic images of breast tissue. We propose a new method for the automatic discrimination of cancerous tissue from normal tissue. The method is based on granulometries, which are size-shape descriptors widely used in mathematical morphology. Applications of granu- lometries lead to distribution functions whose moments are used as input of a neural network classifier to discriminate different tissues. The technique is illustrated with the anal- ysis of some images of breast tissue. KEY WORDS Medical image analysis, breast cancer, granulometry, math- ematical morphology, neural network 1 Introduction This paper is concerned with a medical image analysis problem: namely, the evaluation of abnormalities in mi- croscopic images of breast tissue. The aim of this study is to assist the pathologist in the diagnosis of breast cancer. Breast cancer is the most common form of cancer among women. The diagnosis is usually confirmed by the pathol- ogist by subjective evaluation of tissue samples. An auto- matic system for the analysis of tissue samples would be of great help. Objective analysis of microscopic images of cells and tissues has been a goal of human pathology and cytology since long time. Many studies have been performed to identify and analyze isolated cells according to features measured by image analysis systems, such cell size, shape, texture [1, 2, 3]. The histology encompass the cytology, because the diagnosis of the pathologist is based on the judgment of the entire tissue, i.e. the cells are in their right context and it is possible to observe both the structure of the tissue and the cells themselves. For the automatic discrimination of entire tissue samples, it is therefore necessary to develop a new classification method, based on the observation on how the pathologists judge tissue samples. The most common type of breast cancer is ductal car- cinoma; it originates in the cells of the ducts. Cancer that begins in the lobes or lobules is called lobular carcinoma, which is the second most common type of breast cancer. As described by pathologists, the cells in normal breast tissue are well organized in regular objects, which are all about the same size; the cells do not invade the surrounding tis- sue. The cells in the two types of cancers do not form any kind of regular pattern, because the cells are growing with- out control; the cancerous cells are invading the surround- ing tissue. The observation of the cell organization prompted us to use a morphological method to extract features able to classify the different tissues. Mathematical morphology is well suited for biologi- cal and medical image analysis. In fact, it offers a power- ful tool for extracting image components that are useful in the representation of object shape and size. Our proposed method is based on morphological granulometries. The granulometric methodology models a sieving ac- tion with increasing-sized sieves and the image is classified in accordance with the rate of sieving. The model involves iteratively opening an image by increasing structuring ele- ments so that the image undergoes successive diminution. The decreasing image areas compose a size distribution and the normalized size distribution becomes a probabil- ity distribution function that is characteristic of the image. In practice, the low-order moments of the normalized size distribution are employed as image descriptors. Granulometries have been used successfully for clas- sification of binary textures [4]. Granulometries have been proposed for pattern recognition and in that contest the ter- minology pattern spectrum has been introduced for the nor- malized size distribution [5]. Some works on applications of granulometries to cell images have been recently pub- lished [6, 7]. In this paper we describe the use of morphological granulometries for the analysis and classification of micro- scopic images of breast tissue samples. So, the paper is or- ganized as follow: in Section 2 the theoretical background about mathematical morphology is given. In Section 3, we