Conclusions

In this chapter we firstly reviewed qualitatively the different approaches on breast tissue quantification and classification. We noted that the actual protocol in radiology does not tend to quantify the breast, but to qualitatively describe it. For this reason, we focused on breast tissue classification.

From the review of breast tissue classification approaches we noted that none of them used a segmentation between dense and fatty tissue to classify the breast, and we have proposed a new method that uses this segmentation strategy. To briefly summarize it, once the breast has been segmented from the background and pectoral muscle, a Fuzzy C-Means algorithm is used to segment different tissue types (fatty versus dense) in the mammograms. For each tissue region, morphological and texture features are extracted to characterize the breast tissue. Finally, using a Bayesian approach and obtaining the likelihood estimation by combining both kNN and C$4.5$ classifier results, the mammograms are classified according to BIRADS categories. It should be noted that to avoid bias we have adopted a leave-one-woman-out methodology.

Summarizing the results, we obtained for the MIAS database and individual experts $83\%$ , $80\%$ , and $82\%$ correct classification, which increased to $86\%$ when the classifiers are based on the consensus ground-truth. On the other hand, results based on the DDSM database (a set of $831$ mammograms) showed a performance of $77\%$ correct classification. The strength of the Bayesian classifier might be partially explained by the features that were mainly used by the individual classifiers. The SFS stage of the kNN classifier has a strong tendency to select texture features independently of the distance used for the co-occurrence matrices, whilst most of the selected features for the C$4.5$ classifiers are related to the statistics obtained using a distance equal to $9$ for the co-occurrence matrices.

We exhaustively tested the method using MIAS and DDSM database, showing that our proposal outperforms current works on breast tissue classification using the BIRADS standard. Moreover, we compared the different strategies reviewed, showing also that our proposal obtains better results than the others.