Using neural networks with wavelet transforms for an automated mammographic mass classifier

The objective of this paper is to demonstrate the utility of artificial neural networks, in combination with wavelet transforms for the classification of mammogram mass shapes as round or irregular. The discrete wavelet transform is applied to the radial distance measure of the mass shapes, and scalar-energy features are computed to form the input to the neural network classifier. A two-layer neural network with a backpropagation algorithm was trained to differentiate between two classes of mass shapes that were determined by an expert radiologist. The neural network was trained on the wavelet-based feature vectors extracted from the mammogram mass shapes for both manually and automated-segmented data. The performance of the classification system was studied using receiver operating characteristic analysis. It has been shown that the Haar mother wavelet extracted a set of features from the automated-segmented data to obtain sensitivity and specificity of 100%.