International Journal of Computer Applications (0975 – 8887) Volume 82 – No2, November 2013 44 Performance Analysis of Image Segmentation Methods for the Detection of Masses in Mammograms Zainul Abdin Jaffery Dept. of Electrical Engineering Jamia Millia Islamia New Delhi, India Zaheeruddin Dept. of Electrical Engineering Jamia Millia Islamia New Delhi, India Laxman Singh Dept. of Electrical Engineering Jamia Millia Islamia New Delhi, India ABSTRACT Detection and quantification of breast cancer is a very critical step in mammograms and therefore, needs an accurate and standard technique for breast tumor segmentation. In the last four decades, a number of algorithms have been published in the literature. Each one has their own merits and demerits. The aim of this paper is to make a comparative analysis of the most promising methods, namely fuzzy c-means (FCM), k- means (KM), marker controlled watershed segmentation (MCWS) and region growing (RG), for the detection and segmentation of masses in mammographic images on real data obtained from Metro Hospital. Robustness of the methods is demonstrated by validating their quantitative results with expert manual data. It is observed that the RG gives better results compared to three other methods. Keywords Breast cancer, mathematical morphology, marker controlled watershed segmentation, region growing. 1. INTRODUCTION Cancer is one of the leading causes of human mortality in the world. The most common type of cancer in women is the breast cancer. Early detection and diagnosis leads to the successful treatment and thus play the key role in controlling the breast cancer deaths. Hence, it is essential for the women of age group 30-40 years to have regular screening every year. Currently, X-ray mammography is considered to be the most simple and reliable imaging method for the early detection of breast cancer [1]. Presence of masses or micro-calcification clusters on mammograms is considered to be preliminary indicators for early stage breast cancer. To determine the tumor area, in most of the hospitals, a radiologist performs the diagnosis of breast tumor manually on mammographic images. Visual examination of large volume of mammograms and shortage of experienced radiologists makes the process error prone and time consuming. Computer-aided diagnosis (CAD) system may help radiologist and doctors in reliable and precise diagnosis of breast cancer [2]. Numerous techniques have been developed and proposed as an emerging tool to segment masses from surrounding tissues in digital mammograms. Martins et al. [3] employed the K-means algorithm for mass detection on digitized mammograms and classified them into masses and non-masses through support vector machine using shape and texture descriptors. Dominguez et al. [4] applied three image segmentation methods, K-means, Fuzzy c-means and Possibilistic Fuzzy C- means, for the detection of microcalcification clusters and achieved the better segmentation results with Possibilistic Fuzzy C-means. Kannan et al. [5] proposed a kernel induced fuzzy c-means based hyper tangent function (KFCHF) algorithm for the segmentation of breast cancer from mammographic images. In this work, the objective function of standard fuzzy c-means was modified by replacing original Euclidean distance on feature space using new hyper tangent function and the objective function thus obtained was converged more rapidly. Malek et al. [6] employed seed based region growing and mathematical morphology for the segmentation of microcalcifications in mammograms. Zaheeruddin et al. [7] presented the mean based region growing segmentation (MRGS) and marker controlled watershed segmentation (MCWS) for the segmentation of breast tumor in mammograms. Through the literature survey it is observed that each of the above mentioned algorithms or their modified versions have been used somewhere else but each method on different data sets. Therefore, it is not possible to evaluate their performance directly without implementing all these four algorithms on common data sets. The region growing (RG) and marker controlled watershed segmentation (MCWS) are the region based methods and fuzzy C-means (FCM), K-means (KM) are clustering based methods. In the present work, four prominent methods for the image segmentation, namely, fuzzy C-means (FCM), K-means (KM), marker controlled watershed segmentation (MCWS), and region growing (RG) have been selected and an algorithm is developed for the detection of masses using these methods. This algorithm is tested using a number of mammographic images obtained from a diagnostic centre in New Delhi, India. The performance of these segmentation techniques are evaluated and analyzed. 2. DETECTION METHODOLOGY 2.1 Image Enhancement Mammograms contain noise, uneven illumination and several other artifacts that must be eliminated prior to segmentation. It is very difficult to locate the suspicious tumor areas in mammograms due to minor intensity differences between normal breast regions and anatomical regions. Therefore, pre- processing of mammogram image is essential to eliminate the background noise and enlarge the intensity gap between object and its background. In this work, we have applied morphological open-close reconstruction filter [8] to enhance the contrast between mass site location and background. Mathematical morphology is an effective tool for dealing with various problems related to image analysis and computer vision [9]. Morphological operations, such as, erosion, dilation, opening and closing are used for analyzing and processing of geometric structures based on set theory. Traditional filters, such as opening and closing, is generally used for noise reduction but its use is not widespread in image processing applications because it causes blur effects that poses serious problems in correct segmentation of region of interest (ROI). Morphological reconstruction based opening and closing filter has been proven better in terms of shape preservation than conventional morphological opening and closing operators [10]. Reconstruction by dilation is a morphological procedure that is defined by their respective geodesic dilation operators in binary and gray scale images.