A comprehensive strategy for mammogram image classification using learning classifier systems

Mammography is a well known procedure for breast cancer detection. The traditional mammography process employs manual analysis for detection and diagnosis, which requires professional expertise. However, computerized systems that use feature based classification have been demonstrated to be proficient and reliable but there is scope for improved accuracy. In this paper, we present the first comprehensive strategy to use learning classifier systems (LCSs) for mammogram image classification. We use six types of statistical measures, three different variants of local binary pattern (LBP) technique and ten variants of discrete wavelet transform (DWT) to extract statistical, texture and multiresolution features, respectively. However, the main challenge to apply an LCS in image classification tasks is the large number of extracted feature components that result in a large number of attributes in classifier conditions. Whereas, to evolve generalization in an LCS, a limited number of attributes in classifier conditions are required. We use different encoding schemes based on mapping distances against the large feature components to reduce the number of required attributes in classifier conditions while retaining the unique image characteristics. We develop a novel strategy that deploys various combinations of features and distances to investigate five different types of attributes in classifier conditions: (i) individual statistical features, (ii) individual LBP features, (iii) concatenation of statistical and LBP features, (iv) concatenation of statistical, LBP features and distances based on DWT features, and (v) concatenation of statistical features and distances based on LBP and DWT features. The obtained results indicate that using the precomputed distances in place of the original LBP and DWT features improve the classification accuracy in experiments conducted in this study.