Calculating Texture Features from Mammograms and Evaluating Their Performance in Classifying Clusters of Microcalcifications

In this work, 2432 texture features were calculated from microcalcification clusters presented on 190 images from the Digital Database for Screening Mammography. Mutual information technique was used to rank texture features. Then, an incremental procedure adds top ranked features to the Fisher discriminant analysis to determine the best set of texture features in classifying benign or malignant microcalcification clusters. The result was achieved using 13 texture features (AUC(.632+) = 0.945 +/- 0.019). However, to assure a consistent statistical analysis, at least 30 sample images for each feature added was assumed. The best performance was achieved by a set with 5 texture features (AUC(.632+) = 0.884 +/- 0.025), which is comparable to the ones presented in literature.