The preceding chapters have dealt with low and intermediate-level processing needed to implement a CAE system for breast MRI. This chapter focuses on the classification (high-level processing) of suspicious lesions. It introduces the background and motivation for the automatic classification of the suspicious lesions and reviews the existing methods and features for the classification of the suspicious lesions in breast MRI. It also presents a study whose aim was to determine the most discriminatory subset of features for suspicious lesions in DCE-MRI of the breast. The study shortlists the most discriminatory features for the classification of breast lesions in DCE-MRI that have been reported to date (Section 6.2). It also presents several new features based on the empirical model of contrast enhancement described in Chapter 3, while taking into account the curse of dimensionality, and assessing the classification performance using separate validation data (Section 6.3). The results (presented in Section 6.3.6) suggest that textural and kinetic features are more important than morphometric ones and that the CAE can, indeed, improve the specificity of breast MRI. The core material in this chapter was presented at �Digital Image Computing: Techniques and Applications�, 2009, Melbourne, Australia (Gal et al., 2009a) and was submitted to the Elsevier Artificial Intelligence in Medicine.
Dynamic contrast-enhanced (DCE) MRI is being increasingly used in the clinical setting to help detect and characterise tissue, suspicious for malignancy (Sinha et al., 1997). In an attempt to reduce the subjectiveness of the interpretation, the American College of Radiology developed the BI-RADS (Breast Imaging Reporting and Data System) MRI lexicon (American College of Radiology, 2006) that provides a standard terminology for reporting breast MRI findings (see Appendix D). In particular, the BI-RADS lexicon provides terminology for describing the morphology of a lesion, in addition to its enhancement and kinetic behaviour. A plethora of features for the automatic segmentation of suspicious lesions have been proposed for breast MRI