Abstract

Segmentation is typically the first step in computer-aided-detection (CADe). The second step is false positive reduction which usually involves computing a large number of features with thresholds set by training over excessive data set. The number of false positives can, in principle, be reduced by extensive noise removal and other forms of image enhancement prior to segmentation. However, this can drastically affect the true positive results and their boundaries. We present a post-segmentation method to reduce the number of false positives by using a diffusion scale space. The method is illustrated using Integral Invariant scale space, though this is not a requirement. It is quite general, does not require any prior information, is fast and easy to compute, and gives very encouraging results. Experiments are performed both on intensity mammograms as well as on Volpara® density maps.