Computer-Aided Diagnosis Scheme applying GRNN Neural Network Antonio Vega-Corona, Alexis Marcano, Diego Andina. Universidad de Guanajuato. Mexico Departamento de Señales, Sistemas y Radiocomunicaciones. Universidad Politécnica de Madrid. Madrid, Spain. Email: tono@salamanca.ugto.mx, a.marcano@gc.ssr.upm.es, andina@gc.ssr.upm.es. Abstract— In this paper, we present a Computer Aided Diagno- sis (CAD) system that has been developed for Microcalcifications (MCs) Identification in Digitized Mammography. It combines Multiscale Image Processing and Artificial Neural Networks (ANNs). In mammography diagnosis, Generalized Regression Neural Network (GRNN) is a succesful novelty. A number of relevant features are extracted from Regions of Interest (ROI) to reduce the classifier complexity. We compared our results with a Feed Forward Neural Networks (FFNNs). Our performance results successfully compete with classical FFNNs and present a much faster training phase, overcoming also minima problem. I. I NTRODUCTION In recent studies made by the International Agency for the Cancer Research of World Health Organization, it has been estimated that more than 150.000 women worldwide, die of breast cancer each year [1],[2]. The breast cancer is the most frequent cancer in women, especially in the occidental countries. It’s also a principal cause of mortality each year [3]. The breast cancer early detection using screening mam- mography is currently an available efficient method . In the nationwide screening mammography programs, it’s difficult for radiologists to detect microcalcifications (MCs), since it demands great concentration and time. Therefore, due to a large number of mammographies to be analyzed, the risk of not identifying cancer is great [1]. Many different techniques have been used to developed a CAD system for MCs detection. Recently Nakayama et. al [4] developed a CAD system base in feature extractions and analyzed the shape of the microcalcifications developing a computerized scheme with the potential to detect microcalci- fication clusters with clinically acceptable sensitivity and low false positives. Karssemeijer et. al [5] developed a statistical framework using Bayesian image analysis. Kupinski et al [6] proposed a framework of MCs detection system using Bayesian Neural Network (BNN) optimized as ideal observer. Sonyang et al [7] developed a CAD system based in feature extractions and feature selections applying GRNN and a FFNN. In this paper, we have developed a CAD system that uses a mixture of features extracted from a Region of 1 - - - - - - -- This research has been supported by the National Spanish Research Institution "Comisión Interministerial de Ciencia y Tecnología - CICYT" as part of the project AGL2006-12689/AGR. Interest (ROI) previously diagnosed as teaching element of our CAD system. The paper is organized as follows: In Section 2, Database and features extraction techniques are presented. In Section 3, Feature extraction is presented. In Section 4, the methodology for feature selection and the classifier theory are presented. In section 5, we present and briefly discuss the results of the experimental analysis. In Sections 6, we give the investigation conclusions II. DATABASE AND FEATURE EXTRACTION TECHNIQUES A. Database The images database used in this paper is provided by the Digital Database for Screening Mammography (DDSM) of the University of South Florida [8]. We use 50 cases with MCs pathology (with both malignant and benign lesions). B. Processing of the Region of Interest We focused our analysis on the ROIs, because the relevant information is concentrated in this area. The contrast between normal and malignant tissue usually is present in ROIs, however with an unclear threshold human perception [9]. Like part of the image processing that we make on the ROIs, it’s of interest to detect the limits of the singularities to identify probably calcifications [10]. C. Denoising and Enhanced Based in Gradient-Laplacian Wavelet Transform. In order to analyze and suppress the Gaussian Noise in the ROIs we applied a Gradient method in the transform domain [10]. III. FEATURE EXTRACTION From reconstructed images by the method described in Section 2.3, we analyze the gray levels distribution by means of a self-organizing method in order to segment the MCs areas using non contextual feature vector [11],[12]. For this purpose we defined J=4 levels of wavelet analysis. We obtain four free- noise and enhanced sub-images at each level. Then we mapped the four sub-images to a 4-dimensional feature vector set as follows. {S j f e (x, y): j =1, .., 4} 1 ≤ x ≤ n, 1 ≤ y ≤ m → X (q) = {x (q) j : j =1, .., 4}1 ≤ q ≤ nXm