We have presented a new strategy which is a generic, simple and cost-effective method for false positive reduction. The strategy consists on training a classifier with RoIs representing masses and normal tissue, and using a statistical-based approach to classify a new query RoI as belonging to one of the training set.
We have evaluated two different algorithms: one based on PCA and the other based on 2DPCA. The performance of the system has been evaluated using a leave-one-out methodology and ROC analysis calculated at different ratios of RoIs with masses and RoIs depicting normal tissue. The obtained results demonstrate that, for false positive reduction, the 2DPCA approach outperforms the traditional PCA.
Moreover, we have integrated the approach into the algorithm developed in the previous chapter. The performance of this integration has been evaluated using FROC and ROC analysis, obtaining promising results. Moreover, we have tested the system using the same and different databases for training and testing. We noted that when using different databases the number of false positives as well as the performance of the system decreased, particularly in the smaller masses. This is probably due to the fact we trained the system with a larger number of smaller masses compared to the ones found in the MIAS database, which are slightly larger in size.