Resumen de False positive reduction in detection problems
It is important for detection systems to obtain high detection rates. However, when these systems are configured to increase the detection rate (Dr) the false positive rate (FPr) also increases. In practice, due to the high frequency of negative events, even low false positive rates produce an unacceptably high number of false positive detections. Therefore, methods for reducing this amount of false positives are needed. Since a detection system entails several steps until the decision is made, the improvement of the detector by reducing the FPr can be achieved at different stages. On the one hand, in this thesis an image classification step is applied prior to detection. We show that this step can help reduce the final FPr obtained by the detector. This pre-detection classification step is applied to mammograms classifying them according to the parenchymal densities specified in the BI-RADS (Breast Imaging Reporting and Data System). As a result, a novel hierarchical procedure based on weighted classifiers and texture features has been proposed for breast parenchymal classification. The proposed approach has been tested using 10FCV (10-fold cross-validation) and LOOCV (leave-one-out cross-validation) with the public mini-MIAS database and a large FFDM (Full Field Digital Mammography) database from local hospitals. The obtained results improve upon previously reported results. Moreover, a breast CADe (Computer Aided Detection) system has been developed to incorporate breast parenchymal density information. The results show that the proposed classification helps to adjust the parameters of the CADe algorithms and decrease the false positive rate. On the other hand, and also with the objective of reducing false positive in detector systems, this thesis focuses on the widely-used cascade detector. A sample selection method for training cascade detectors is proposed to achieve good detection/false positive ratios. The method is based on the selection of the most informative false positive samples generated in one stage. Then, these samples are used to feed the next stage. The proposed cascade framework with sample selection was compared with other cascade detectors using different databases and feature sets. The effectiveness of the method was assessed with the average partial AUC (Area Under the Curve) from the ROCs (Receiver Operating Curves) obtained with 10FCV. The results show that the proposed cascade detector with sample selection obtains on average better results.
