Multi-fractal dimension features by enhancing and segmenting mammogram images of breast cancer

Abstract

The common malignancy which causes deaths in women is breast cancer. Early detection of breast cancer using mammographic image can help in reducing the mortality rate and the probability of recurrence. Through mammographic examination, breast lesions can be detected and classified. Breast lesions can be detected using many popular tools such as Magnetic Resonance Imaging (MRI), ultrasonography, and mammography. Although mammography is very useful in the diagnosis of breast cancer, the pattern similarities between normal and pathologic cases makes the process of diagnosis difficult. Therefore, in this thesis Computer Aided Diagnosing (CAD) systems have been developed to help doctors and technicians in detecting lesions. The thesis aims to increase the accuracy of diagnosing breast cancer for optimal classification of cancer. It is achieved using Machine Learning (ML) and image processing techniques on mammogram images. This thesis also proposes an improvement of an automated extraction of powerful texture sign for classification by enhancing and segmenting the breast cancer mammogram images. The proposed CAD system consists of five stages namely pre-processing, segmentation, feature extraction, feature selection, and classification. First stage is pre-processing that is used for noise reduction due to noises in mammogram image. Therefore, based on the frequency domain this thesis employed wavelet transform to enhance mammogram images in pre-processing stage for two purposes which is to highlight the border of mammogram images for segmentation stage, and to enhance the region of interest (ROI) using adaptive threshold in the mammogram images for feature extraction purpose. Second stage is segmentation process to identify ROI in mammogram images. It is a difficult task because of several landmarks such as breast boundary and artifacts as well as pectoral muscle in Medio-Lateral Oblique (MLO). Thus, this thesis presents an automatic segmentation algorithm based on new thresholding combined with image processing techniques. Experimental results demonstrate that the proposed model increases segmentation accuracy of the ROI from breast background, landmarks, and pectoral muscle. Third stage is feature extraction where enhancement model based on fractal dimension is proposed to derive significant mammogram image texture features. Based on the proposed, model a powerful texture sign for classification are extracted. Fourth stage is feature selection where Genetic Algorithm (GA) technique has been used as a feature selection technique to select the important features. In last classification stage, Artificial Neural Network (ANN) technique has been used to differentiate between Benign and Malignant classes of cancer using the most relevant texture feature. As a conclusion, classification accuracy, sensitivity, and specificity obtained by the proposed CAD system are improved in comparison to previous studies. This thesis has practical contribution in identification of breast cancer using mammogram images and better classification accuracy of benign and malign lesions using ML and image processing techniques.