Comparative Study of a Shape-Based and a Texture-Based Feature Extraction Technique for Mass Classification in Digital Mammograms

This paper presents a comparison of a shape-based and texture-based feature extraction technique. The method use shape features extracted by curvelet transform and texture features extracted by local binary pattern for classification of normal and abnormal mammograms and further classification of abnormal mammograms into cancerous (malignant) or non-cancerous (benign) masses. In this experiment, normal and abnormal (mass only) mammograms were obtained from Mammographic image analysis Society (MIAS). The breast images were pre-processed and segmented using histogram normalization and active contour, respectively. Then K-Nearest neighbor classifier was used for classification to evaluate the features extracted by curvelet transform and local binary pattern. The classifier produced a classification accuracy of 73.3% using curvelet transform features and 83.3% using local binary pattern for normal/abnormal classification phase. For benign/malignant classification phase, accuracy of 72% was obtained using curvelet transform features and 84% using local binary pattern features. The experimental results suggest that the texture-based feature extraction technique yield better classification results than the shape-based feature extraction technique.