Segmentation of masses in digital mammograms

This paper suggests a methodology for segmentation of masses in digital mammograms. The masses are distinguished from other breast tissue by its homogeneous and differentiated density in relation to other breast tissues. The segmentation strategy is based on the assessment of density using multiscale wavelet transform. The density data obtained by processing with wavelet are used to train multilayer perceptron network (MLP) with one hidden layer with error back-propagation algorithm. After the training phase, any mammography with possible masses is then submitted to a trained neural network. The image resulting from processing handled by the neural network has evidenced the relevant characteristics of the original image. The characteristics not relevant were minimized with respect to density. The processed image then serves to provide contours of possible masses in the original image using an automated thresholding criterion. A set of five images was used in the training phase. The trained network was used to detect masses in 19 images (not used for training) that were previously classified by an expert. The TPR (sensitivity) measured was 68.2%, the FPR measurement was 8.7%.