This paper proposes a CAD system for the automatic detection and classification of microcalcifications in digitized mammograms. Features extracted from both the segmented MCs and their surrounding tissues, using GLCM and GLRLM texture features matrices, in addition to LTEM, non Shannon entropy and the classical FOS features are used. A novel approach, for feature selection is created and used to reduce the extracted features to their best informative subset. The performance of three classifiers is invistegated. One is the Linear Discriminant Analysis (LDA) with cross-validation, the second is a Multilayer Perceptron neural network (MLP), and the third is GRNN, a neural network that employs a base of radial functions for functional approximation. MCs are classified into benign or malignant. The accuracy of their performances with the full set of extracted features and the best subset of features, are evaluated and compared using the mammographic data from the Mammographic Image Analysis Society (MIAS) database. A training accuracy of 100% and a testing and validation accuracies of 100 % for MLP and LDA and 97.80 % for GRNN are achieved, outperforming many of previous CAD systems results.