An objective measure for assembling databases used to train and test mammogram CAD algorithms

We calculate a measure of lesion subtlety that is derived using both lesion and parenchymal feature salience. Previously, this measure was shown to correlate well with radiologists´ localization and discrimination of true positive and true negative regions-of-interest. Based upon conspicuous spatial frequency features, an image difficulty rating is obtained that is used to cluster groups of images to establish test sets of known difficulty. Using 100 cases, obtained from the University of Central Florida´s Digital Database for Screening Mammography (DDSM), we show that test sets can be objectively constructed. Performance is assessed using a linear observer model, which correlates test set difficulty to detection performance