Two-Level Breast Cancer Staging Diagnosis
Sahar Bayoumi, Sanaa Ghouzali, Souad Larabi-Marie-Sainte, Hanaa Ahmed Kamel · 2021
According to the World Health Organization, the majority of women are mainly affected by breast cancer, which yearly impacts 2.1 million women in the world. Also, breast cancer is known as the cause of the most significant number of cancer-related deaths among women, where, in 2018, approximately 15% of all women cancer deaths are caused by breast cancer. Screening and early diagnosing are two main strategies to reduce the mortality rate and the cost of treatment by detecting the disease in early stages. The most common screening tool is mammography, which has reduced breast cancer mortality by approximately 20%. Mammography is recommended regularly for women aged above 40 years old. However, two different radiologists should analyse the mammogram image to achieve high accuracy of correct detection and diagnosis of breast cancer. Computer-aided detection/diagnosis (CAD) received extensive research on mammography to support specialist decisions. CAD comprises different phases, including region of interest segmentation, feature extraction and classification. To solve the problem of breast cancer classification, a wide range of machine learning methods has been proposed. Recently, deep learning approaches have gained considerable attention as they have proven to enhance detection results. However, most of the studies focused on differentiating between normal and abnormal cases, or between malignant and benign cases. To the best of one&s;s knowledge, the detection of breast cancer stage/severity has not yet been addressed. Our research aims at providing a two-level algorithm for detecting and diagnosing breast cancer in mammogram images. The first level is about classifying abnormal breast tissues into benign or malignant using a support vector machine (SVM) and K-nearest neighbors (KNN) classifiers. Before classification, images are preprocessed to segment the breast tissue and converted into a vector of extracted features. It has been shown in the literature that breast cancer classification can be enhanced when features are extracted using multi-resolution analysis such as wavelet, curvelet and contourlet transforms. The choice of using SVM and KNN for the initial level classification is due to their low complexity and promising results, especially when fed with pertinent features such as the wavelet coefficients and texture-based features. The second-level classification aims to specify the severity of malignancy, such as early or advanced stage using the same classifiers based on pertinent features. The efficiency of the proposed approach is validated using the mammogram images from Mammographic Image Analysis Society database, which contains 115 images tagged as “abnormal” covering 51 benign samples and 64 malignant samples.