Segmentation of Pectoral Muscle in Mammograms Using Granular Computing

1. Introduction

For decades, breast cancer is considered as number one potentially lethal disease reported among women worldwide. According to Union health ministry, government of India, breast cancer occurred is about 25.8 per 100,000 women and mortality reported is 12.7 per 100,000 women (Malvia et al 2017). Early detection of breast cancer is crucial in reducing the mortality rate (WHO, 2017) and increases treatment options. Mammography is the most powerful and successful breast screening tool that is used to detect cancer in the early stage. A breast mammogram contains different parts: mainly the breast region, pectoral muscle, high-density label, low-density label and background (Figure 1(a)). The computer-aided classification of mammograms into normal and abnormal is an emerging research area in the field of medical image processing. So far, the identification of malignancy in the breast region in mammograms, ignoring the pectoral muscle, has been attempted. However, a recent study states that the presence of masses in the pectoral region indicates that it could also be malignant (Figure 1(b, c, d)). Hence, it is important to analyze both the pectoral muscle and the breast region. In this regard, the idea of this article is to segment the pectoral region to study the presence of malignancy.

Figure 1.

(a) Different parts of mammogram; (b) Abnormality marked in the pectoral muscle region

The segmentation of the pectoral muscle in mammograms is a challenging task because the mammogram images have low contrast. Further, the appearance of the pectoral muscle varies in size, shape, and orientation (left and right), having unclear boundaries between the pectoral region and the breast region (Mustra, Grgic & Rangayyan, 2016). In the state-of-the-art, researchers have attempted to segment the breast region by suppressing the pectoral muscle. However, in this paper, the effort is to segment the pectoral muscle for further analysis using granular computing and the ant colonization approach. Granular computing is used by researchers for exploring the masses in mammograms with some different statistics (Roselin & Thangavel, 2012) but in the proposed work, only pixel intensity values and mean statistics are considered for pectoral muscle segmentation. This step helped in eliminating additional enhancement techniques and reduced the processing time by removing the artifacts and labels present in the background region automatically. Ant colony optimization algorithm adopted in the proposed paper is slightly different from the ant colony optimization algorithm used in the state of art (Karnan, Thangavel & Ezhilarasu, 2008) which is discussed in section 3. The proposed method was tested on the benchmark mini MIAS dataset.

The paper is organized as follows: the related research work is presented in section 2; the materials and methods are discussed in section 3; section 4 shows the experimental results and discussion and section 5 draws conclusions with future avenues.

2. Related Works

Many researchers have proposed methods with regard to pectoral muscle boundary identification and removal. Existing studies on pectoral muscle segmentation focused on different approaches to accurately segment the pectoral muscle, among which the most popular ones are region growing, polynomial fitting, straight-line estimation and thresholding methods.