A New Proposal for Automatic Breast Density Classification

The first step in our approach is the segmentation of the breast profile, which is done using the algorithm explained in Appendix . As explained, this segmentation provides a minor loss of skin-line pixels in the breast area. In this case, these pixels are also deemed not to be relevant for tissue estimation and, in addition, the relative number of potentially affected pixels is small. The second row in Figure  shows examples of the breast segmentation.

Therefore, once the breast is segmented, our approach will go beyond the use of histogram information obtaining a set of features for characterizing the mammogram. From the reviewed literature, we can distinguish two related strategies. Bovis and Singh [17] extracted a set of features using the global breast area, hence assuming that the breast is composed of a single texture. As shown in the results of Figure (a,b) (and in the mammograms shown in Figure ), in many cases this is hard to justify. On the other hand, Karssemeijer [83], and subsequently Blot and Zwiggelaar [14], divided the breast into different regions according to the distance between pixels and the skin-line, as is shown in Figure (c). The main idea for such approach is the assumption that a strong correlation between tissue density and distance to the skin line will exist. However, note from Figure  (and again in the mammograms shown in Figure ) that using this strategy it seems that tissue with the same appearance (texture) is divided over different regions, as well as tissues with different appearance are merged in the same region. In contrast with these approaches, our proposal is based on the segmentation of the breast in order to group those pixels with similar tissue appearance, as is shown in Figure (d). Subsequently, extracting a set of features from each region, different classifiers are trained and tested. A quantitative evaluation of these strategies is provided in Section .

Figure 3.2: Three different strategies for dividing a mammogram (a) into regions: (b) whole breast area, (c) based on the distance between pixels and the skin-line [14,83], and (d) based on clustering pixels with similar appearance.

(a)	(b)	(c)	(d)