Abstract
Masses are one of the early symptoms of breast cancer and mammography is an effective methodology for the early detection of masses. For mass detection, the segmentation of mammograms generates regions of interest (ROIs) that represent not only mass areas but normal tissues as well leading to false positives. This results in the problem of false positive reduction (i.e. classifying ROIs as masses and normal tissues). Texture properties of masses provide powerful discriminative information and the texture micropatterns can be effectively represented using Gabor filters with different scales and orientations. Though a local texture descriptor based on Gabor filter bank represents multiscale and multidirectional texture micropatterns of masses effectively, it involves a bulk of redundant features, increasing the dimensionality of the feature space and reducing the classification performance. This poses a challenge to even some of the most modern classification algorithms such as Support Vector Machine (SVM). To tackle this issue, the effectiveness of two state-of-the-art feature subset selection algorithms has been investigated in this study. The proposed approach has been evaluated on 768 (256 masses and 512 normal) ROIs extracted from the DDSM database. The best result (Az = 0.99 +/- 0.01) was obtained using a Gabor filter bank with 8 orientations and 5 scales and RIOs of size 128x128. Comparison with state-of-the-art methods reveals the superiority of the proposed method.