Breast Profile Segmentation

Although segmentation of the breast from the background is not the main goal of this review, this kind of segmentation is a fundamental step in mammogram image analysis because the techniques which will be discussed in Sections  and  are only applied to the breast area. For this reason, we briefly review different approaches to breast profile segmentation found in the literature.

The aim of breast profile segmentation is to separate the breast from other objects in the mammogram with a minimum loss of breast tissue. In general two independent steps are performed. The first one aims to segment the background and annotations from the whole breast area, while the second one involves separating the pectoral muscle (when present) from the rest of the breast area. Approaches to the breast-background segmentation range from simple histogram thresholding followed by smoothing [69,100] to polynomial modeling [27], or, more recently, active contour approaches [49]. Typical strategies to segment the pectoral muscle have been based on straight line estimation using a Hough transform [84,97] or direct detection using Gabor filters as edge detectors [50].

Figure 2.1: Breast profile segmentation of two mammograms using the algorithm explained in Appendix , which is the used in this work.

In this work, a new approach has been designed for such task. Firstly, an automatic thresholding algorithm is used to separate the area composed of the breast and the pectoral muscle from the background of the image. Subsequently, a region growing algorithm allows to locate the muscle and extract it from the breast. The result of this approach is a visually correct segmentation of the breast, although some pixels belonging to the skin-line are misclassified as background. However, this set of pixels is not relevant in posterior steps. Figure  shows two examples of the performance of these algorithms. This approach is described in detail in Appendix .

To avoid the misclassifications of the pixels with low grey-level value located near to the skin-line a new algorithm has also been designed applying edge detection and scale space concepts. Thus, the main edges of the image are correctly located and subsequent post-processing is used to isolate the skin-line. Figure  shows the performance of this algorithm in combination with the pectoral muscle segmentation proposed by Ferrari et al. [50]. Appendix  gives more details of this skin-line detection algorithm.

Figure 2.2: Breast profile segmentation of two mammograms using the algorithm described in Appendix . This algorithm segments all pixels in the breast, although the boundary seems far away of it.