Background texture extraction for the classification of mammographic parenchymal patterns

We have developed an approach to the separation of background texture and structures in images. The developed approach is based on the statistical difference between local and median co-occurrence matrices. It is our assertion that the classification of mammographic parenchymal patterns can be improved if anatomical structures can be removed from the image and the classification is based only on the background texture information. We compare the results of the classification between original images and images composed of their reconstructed background texture. 265 mammograms from the MIAS database [1] have been used for our experiment and the classification of the parenchymal patterns is based on Karssemeijer’s model [2].