Effects of Dimension Reduction in Mammograms Classification

Abstract

Breast cancer is the most common type of cancer among women and causing deaths in women. In this paper, a CAD system is presented to investigate effects of dimension reduction for classifying mammograms. Proposed system consists of preprocessing, feature extraction, dimension reduction and classification steps. Multiscale top-hat transform is used to enhance mammograms and to remove noise. First order and second order textural features are extracted from enhanced mammograms. Principal component analysis (PCA) is used for dimension reduction. Two multilayer perceptron neural networks (MLP) are used to classify mammograms as normal or abnormal. All twenty features (without PCA) and selected seven features by PCA are applied each of two classifiers. First MLP classifier with all features achieved accuracy of 79,4%. Second MLP classifier with selected features by PCA achieved accuracy of 91,1%. PCA feature dimension reduction improved the classification performance, increasing accuracy value from 79,4% to 91,1%.