Introduction

Mammographic Computer Aided Diagnosis (CAD) systems are being developed to assist radiologists in the evaluation of mammographic images [13,52]. However, recent studies [75], as well as the results presented in the previous chapter, have shown that the sensitivity of these systems is significantly decreased as the density of the breast increases, while the specificity of the systems remains relatively constant. In addition, it is well-known that there is a strong positive correlation between breast parenchymal density in mammograms and the risk to develop breast cancer [193]. As Taylor [173] suggested, the development of automatic methods for classification of breast tissue are justified, at least, by two factors:

• To permit better use of the time and skills of expert radiologists by allowing the difficult mammograms to be examined by the most experienced readers.
• To increase the scope for computer-aided detection of abnormalities by concentrating on the easier (fatty) mammograms.

In this thesis, we concentrated on the second factor, not only to detect breast cancer in ``easy" mammograms, but also to establish an optimal strategy to look for mammographic abnormalities, as will be shown in Chapter .

The origins of breast density classification are the work of Wolfe [193], who showed the relationship between mammographic parenchymal patterns and the risk of developing breast cancer, classifying the parenchymal patterns in four categories. Since the discovery of this relationship, automated parenchymal pattern classification has been investigated, as is explained in the next section. One of the main variations in those publications is that they classified the breast density using various numbers of density categories/scales [123]. However, the American College of Radiology (ACR) Breast Imaging Reporting And Data System (BIRADS) [2] is becoming a standard on the assessment of mammographic images, not only in the US, but world-wide. In this standard, breasts are classified in four categories according to their density (see Figure  for mammogram examples):

• BIRADS I: the breast is almost entirely fatty.
• BIRADS II: there is some fibrogandular tissue.
• BIRADS III: the breast is heterogeneously dense.
• BIRADS IV: the breast is extremely dense.

In this chapter we review different approaches to automatically classify the breast according to their internal tissue and, moreover, we present a new approach for classifying them according to BIRADS categories. The proposed approach assumes that mammograms belonging to different BIRADS categories are represented by tissue with different texture features. One of the novel aspects of the proposal is the use of an initial fatty versus dense tissue segmentation in order to group pixels with similar tissue characteristics. Subsequently, extracting and comparing texture features from each cluster, the system learns how to differentiate mammograms belonging to each class.

The remainder of this chapter is structured as follows: Section  shows a survey of the methods found in literature, explaining the main strategies and showing the key points of each method. Section  describes the proposed segmentation and classification method. Experimental results indicating the validity of the developed approach are presented in Section , where a quantitative comparison among the reviewed strategies is also done. Finally, discussion and conclusions are given in Section .