Improving GRNNs in CAD Systems Fulgencio S. Buendía Buendía 1 , J. Miguel Barrón-Adame 2 , Antonio Vega-Corona 2 , and Diego Andina 1 1 Universidad Politécnica de Madrid Departamento de Señales, Sistemas y Radiocomunicaciones, E.T.S.I. Telecomunicación, Madrid, Spain {wac,diego}@gc.ssr.upm.es 2 Universidad de Guanajuato F.I.M.E.E., Guanajuato, México tono@salamanca.ugto.mx, miguel@gc.ssr.upm.es Abstract. Different Computer Aided Diagnosis (CAD) systems have been recently developed to detect microcalcifications (MCs) in digital- ized mammography, among other techniques, applying General Regres- sion Neural Networks (GRNNs), or Blind Signal Separation techniques. The main problem of GRNNs to achieve an optimal classification perfor- mance, is fitting the kernel parameters (KPs). In this paper we present two novel algorithms to fit the KPs, that have been successfully applied in our CAD system achieving an improvement in the classification rates. Important remarks about the application of Gradient Algorithms (GR- DAs) are assessed. We make a brief introduction to our CAD system comparing it to other architectures designed to detect MCs. 1 Introduction Breast cancer is a major cause of death among women; several researches have been presented to develop CAD systems capable to detect MCs in digitalized mammographies [6–9], that is an early symptom of breast cancer. Next figure shows the overall architecture of our CAD system: Image Segmentation Feature Extraction (Feature-Based in Wavelet and Gray Levels) Classifier Based in GRNN Feature Selection Based in SFS-GRNN Database Fig. 1. Block diagram of our CAD system. C.G. Puntonet and A. Prieto (Eds.): ICA 2004, LNCS 3195, pp. 160–167, 2004. c  Springer-Verlag Berlin Heidelberg 2004