ABSTRACT
However, an exploratory analysis on the images used in this work showed that a percentile-based technique aimed to reduce the number of false positives is discouraged due to the fact that one would require percentile values well above the 2.5% used in (Lai et al. 1989) (in our work, this threshold would have to be comprised between 2.5 and 10% for most of our images and, in some situations, and it would have to be up to 30%). Besides, the remaining techniques employed in (Lai et al. 1989) are based on some image heuristics whose relations may be difficult to define. Therefore, in order to reduce the number of false positives, we propose an automatic method based on a neural network (NN), which is trained using information extracted from the previously generated cross-correlation images. The use of NN in CAD systems is not new; it has been studied by several authors with very good results (Sahiner et al. 1996, Kaker et al. 1995), but using different approaches. In this work, we also assess if the histogram equalization of the mammographic images can improve the performance of the overall CAD method.
