Automatic breast mass detection in mammograms by mass type-dependent sparsity

Abstract

Breast cancer is the most common type of cancer in women worldwide. To detect breast cancer at an early stage for efficient treatment, computer-aided detection (CAD) systems are actively being developed. Of the stages composing a CAD system, classification stage performs the core function in differentiating the true breast masses from the normal tissues. In order to successively improve detection rate of breast masses using CAD systems, therefore, it is highly recommended to develop a well-organized classifier for the classification stage. Sparse representation based classification (SRC), attracting considerable research interest in the field of signal processing, is a classification algorithm exploiting the sparsity of a given signal. When classifying breast masses via SRC, it is expected to effectively represent the characteristics of the input breast masses as few training samples holding very similar characteristics with the input. If sparsity increases, it is possible to represent the input with fewer training samples, namely, to improve the classification performance. Although the goal of CAD systems is to classify a breast mass or a normal tissue, breast masses present various types of margin and shape. This problem increases diversity in the dictionary, and degrades sparsity of the dictionary. To cope with the problem, in this paper, we divide and conquer the mass classification problem including various margins in order to make each dictionary more sparse. We construct one dictionary using one type of masses. Thus, the number of dictionaries equals to the number of mass margins used. This solution maximizes the characteristics of a breast mass contrasting with normal tissues. Comparative experiments have been conducted on public mammogram data set and the clinical data set provided by a private hospital. Our results show that the proposed method improves the sparsity of the dictionaries divided according to its type of mass margins, and outperforms th...