Abstract
The visual similarity between normal breast tissues and abnormal lesions in digital mammogram images makes computer-aided diagnosis of breast cancer using automatically detected features a highly error-prone task. Our contribution in this paper is a novel feature reduction framework for selecting the most discriminative features that achieves both efficiency and classification accuracy. Our approach applies five individual feature-ranking methods including Fisher score, minimum redundancy-maximum relevance, relief-f, sequential forward feature selection, and genetic algorithm for sorting the extracted features and selecting the features with highest ranking to setup a classifier. Our method achieves an accuracy of 94.27% and a sensitivity of 98.36% with a specificity of 99.27% on a set of 1,100 mammogram patches taken from image retrieval in medical applications database using a neural network classifier, which competes with state-of-the-art classification accuracy 93.11%. Furthermore, we demonstrate that only 49 out of the 119 extracted features are sufficient to achieve the reported accuracy of normal vs. abnormal classification.