Influence of Contrast Enhancement to Breast Density Classification by Using Sigmoid Function

Abstract

The classification of breast density is very subjective even for the experts, the categories 2 and 3, in many cases are confused. Thus, the objective of this study is to evaluate the influence of the sigmoid function in the density classification of images with lesions by using texture attributes. It was used 28 images with lesion, 19 belong to the category 2 (P2 - partially fat) and 9 to category 3 (P3 - dense). The sigmoid function was implemented and applied to all images to contrast windowing. A set of 14 Haralick descriptors were implemented. After the attributes extraction step it used the clustering technique K-Means to classify the category images of breast density 2 and 3. Seven of the 14 Haralick descriptors (energy/ uniformity, contrast, variance/homogeneity, average sum, variance of the sum, entropy difference and maximum correlation coefficient) showed higher success rate when used images processed with sigmoid function. However, five attributes (correlation, entropy sum, difference variance, correlation information measured 1 and 2) presented classification results below those that results by using the original images, and two attributes (inverse difference moment and entropy) obtained the same results, for classification of both images (images with sigmoid function and original images). The attributes combination used to classify images with sigmoid function were better and the combination that had the best classification accuracy rate was the contrast and variance attributes. The use of the sigmoid function directly influenced the results of classification in the categories 2 and 3, however, when it was used by only one attribute in the classification, not all attributes showed great correct response rates, as happened to the results obtained using attributes combination.