CLASSIFICATION OF MAMMOGRAPHIC MASSES USING RADIAL BASIS FUNCTIONS AND SIMULATED ANNEALING WITH SHAPE, ACUTANCE, AND TEXTURE FEATURES

摘要

We investigated the use of a classifier based upon nonlinear and combinational optimization techniques (RBF - radial basis functions and simulated annealing) to classify mammographic masses as malignant or benign. The RBF-simulated annealing network was trained with measures of shape, acutance, and texture. To demonstrate the effectiveness of the classifier in identifying malignant tumors and benign masses in mammograms, the network was trained with the three most effective features of fractional concavity, acutance, and sum entropy. The performance of RBF-simulated annealing was compared with linear discriminant analysis (LDA) in terms of the area (A_z) under the receiver operating characteristics (ROC) curve. The best result obtained with RBF-simulated annealing was A_z = 0.96, which compares well with the result obtained with LDA (A_z = 0.99) using the same features.