Improvement in Ann Performance by the Selection of the Best Texture Features from Breast Masses in Mammography Images

Abstract

Artificial Neural Network (ANN) are frequentely used as classifiers in Computer Aided-Diagnosis (CAD) schemes. The classifier performance depends on the features extraction to represent each class. In this work, 14 features were extracted from mammographic images by using Haralick texture technique. Regions of interest (ROI) containing mass and normal regions were used. By using Gaussian distribution of features into classes, identifing the distance between classes for each feature was possible select the bests. The implemented Multi Layer Perceptron (MLP) and Self-Organizing Map (SOM) classifiers, tested with the 14 selected Haralick texture features, yields a rate of 86.67% and 81.66% of efficacy, respectively, in classifying the cases of suspicious nodules. However, after performing the Gaussian distribution analysis, the 6 best texture features were selected. Thus, tests with the same MLP and SOM classifiers using this new restrict features set have resulted in na efficacy of 91.50% for MLP and 85.83% for SOM in the cases of suspicious nodules, which means an improvement of 4–5% in the classifier performance.