Combining the Bayesian Pattern Matching and the False Positive Reduction Step

In this section we evaluate the results of the combination of both developed algorithms: the Bayesian template matching algorithm developed in Chapter  and the false positive reduction approach developed in this chapter. Thus, for training the algorithms two different subsets are necessary: one to construct the template for finding the suspicious regions and another one to discard the suspicious regions actually being normal tissue. It is important to note that the first subset contains only RoIs depicting masses, while the second one contains different kind of RoIs. Moreover, we only use the 2DPCA approach for false positive reduction since we have demonstrated that outperforms PCA.

Similar to Section  the evaluation is done using a leave-one-out methodology and Receiver Operating Characteristics (ROC) analysis. In the leave-one-out methodology, each mammogram is analyzed using a model created with the rest of RoIs not belonging to this mammogram, and this procedure is repeated until all mammograms have been used as a query image. The evaluation is done here using the MIAS database.