Comparison of artificial neural networks using texture parameters in the recognition of lesions in mammograms digitized

This work proposes to use Radial Basis Function - RBF artificial neural network and Multi-Layer Perceptron MLP with the algorithm cross-validation leave-one-out, to reduce the false-positives of suspicious regions automatically detected by a difference-of-Gaussian filter in mammography. This method was applied to 175 mammograms (one real lesion/image), from the Digital Database for Screening Mammography. Was located and segmented 75.4% of lesions, with 3.55 false-positives/image. In this study, five texture parameters of real lesions and false-positive regions were extracted from a gray-level co-occurrence matrix. These parameters were input of the MLP network, trained with different backpropagation settings, and also input of the RBF network. False-positives were reduced to 1.38 per image, with 0.67 false-negatives per image. Future tests include a greater number of images to enhance the network generalization capacity.