Relevant Features for Classification of Digital Mammogram Images

Abstract

Breast cancer incident in Indonesia reaches 26 per 100,000 women. An early detection of breast cancer is a helpful effort for reaching a successful treatment. Mammography is the best tool for such detection, especially by means of Computer Aided Diagnosis (CAD). The systems of CAD are used to assist the radiologist to determine the benign or malignant abnormalities in the breast. Mammogram image processing system generally consists of mammogram image acquisition, pre-processing, segmentation, feature extraction, feature selection and classification. The features used in feature extraction should be able to represent the characteristics of mammogram image. A feature extraction process uses some texture features based on Gray Level Co-occurrence Matrix (GLCM) and histogram. This study used 60 mammogram images, left and right, from Clinical Oncology Kotabaru Yogyakarta. After passing through the enhancement process, mammogram images were extracted with 11 features of GLCM and histogram. The result then showed that the texture features could be used for the mammogram image feature extraction, but not all of the features were relevant. Thus, for knowing the effects of using irrelevant features, the classification results by using all features and selected features were compared. The highest accuracy was obtained from the selected features reaching at 86.67 %. High accuracy was determined by the relevant features used as input classifier. The selected features here included IDM, ASM, Energy, Contrast, Entropy-based GLCM, Histogram-based Entropy, and Skewness.