Feature extraction for detection of masses in digitized mammograms

Abstract

Breast cancer is one of the major causes of mortality increase to middle aged women, especially in developed countries. X-ray Mammography associated with clinical breast examination is the most effective method for early detection of breast cancer. Masses on mammograms are an important sign in the detection of breast cancer. Mammogram interpretation has been performed by radiologists by visual examination of the films for the presence of abnormalities that can be interpreted as cancerous changes. The computer-aided diagnosis (CAD) will be useful to increase the diagnosis sensitivity of radiologists. The CAD can be categorized into three groups, such as the image enhancement, the detection of suspicious lesions, and the classification of suspicious lesions as benign or malignant. In this thesis, we propose the detection method of mass regions in digitized mammograms using an iterative region growing method (IRGM) and an artificial neural network (ANN). Adaptive histogram equalization is performed to normalize the background and then a segmentation is performed, in which the image is thresholded to select mass candidates and morphologically filtered to smoothen those shapes. Once the mass candidates are decided, the shape-based features are extracted by the IRGM and are used in an ANN classifier. The proposed method, the IRGM, is compared with the conventional texture analysis methods such as the spatial gray-level dependence method (SGLDM) and the gray-level difference method (GLDM). Receiver operation characteristic (ROC) methodology is used to evaluate the classification accuracy. We use 68 mammograms and segment the 56 biopsy-proven masses and 235 normal breast tissues from these data. This data set is used for training and testing the ANN. As a result, the area under the ROC curve reaches 0.93, which corresponds to a true-positive fraction of 90% at a false-positive fraction of 12%.