In recent years, many breast cancer CAD studies have focused on the use of artificial neural
network (ANN) models. ANN models have been developed to predict malignancy among suspicious
breast lesions based upon mammographic and history findings. Most networks for CAD are
based on classic feed-forward, error-backpropagation paradigms, which are trained to minimize mean
squared error (MSE) using a gradient descent technique. In “weight space”, the ANN modifies
a vector of weights, descending down a multi-dimensional error surface in search of the global
minimum in MSE. Once trained, however, these ANNs are often evaluated according to other more
clinically relevant measures of performance from receiver operating characteristic (ROC) analysis.
Such measures include the ROC area index (Az) and the partial area index (0:90Az) corresponding to
the portion of the ROC curve in the high sensitivity range of 0.9–1.0 [9,10]. (More information on
the 0:90Az is provided in the Methods section.)