Abstract:
The early detection of breast cancer is of the outmost importance in order to increase survival rates. Mammography is considered as an effective tool to detect breast cancer, but due to the complexity of breast tissue texture and the observer variability effect, misdiagnosis of breast cancer frequently occurs. This research develops a simple Computer-aided detection (CAD) system in order to improve the accuracy and efficiency of mammogram interpretation. The first step of the proposed algorithm is mammographic image preprocessing which removes unwanted regions from the breast image and enhances image contrast without losing the image information. A region of interest (ROI) was then segmented using thresholding, and quantization based techniques. In mammography, the breast tissue type influences the performance of the detection. Thus, the proposed algorithm automatically characterizes the tissue type as fatty, dense or glandular. The last step of the proposed CAD system, aims at distinguishing normal from abnormal breast tissue, and subsequently differentiates benign from malignant lesions. In this thesis work, the classification tasks were based on texture features and supported vector machine (SVM). Features were extracted by using conventional well-known features (Haralick features) as well as new, less-known features (segmented fractal texture analysis SFTA). To obtain good classification performances, optimal features were selected and redundant features were removed. With the aim of showing the robustness of our approach, tests were performed using the well-known Mammographic Image Analysis Society (MIAS) database which contains annotations provided by radiologists. For tissue characterization, three approaches of features extractions were investigated; the best accuracy for distinguishing fatty tissue from non-fatty was obtained using combined Haralick and SFTA features (87% accuracy). Moreover, the best accuracy for differentiating between dense and glandular tissue was obtained using SFTA features (78 % accuracy). The SFTA features proved their superiority for the classification of normal vs abnormal lesion, and benign vs malignant lesion, with accuracies of 75.5% and 74.5% respectively.
