Radial basis functions-simulated annealing classification of mammographic calcifications

We investigated several approaches to classify mammographic calcifications as malignant or benign: a supervised classifier, multi-layer perceptron (MLP), a supervised and unsupervised classifier, a classifier based upon adaptive resonance theory with linear discriminant analysis (ART2LDA), and a classifier based upon nonlinear and combinational optimization techniques: RBF (radial basis functions)-simulated annealing. The classifiers were trained using shape factors extracted from 143 mammographic calcifications (79 malignant and 64 benign), adopting the leave-one-cut procedure. The classifiers' performance was compared in terms of the area under the ROC curve. The best result of 0.97 was obtained with RBF-simulated annealing, which was significantly better than the results obtained with MLP and ART2LDA, which were, respectively, 0.70 and 0.71.