Applications of Intelligent Optimization in Biology and Medicine

Abstract

Early prediction of breast density is clinically significant as there is an association between the risk of breast cancer development and breast density. In the present work, the performance of two computer aided diagnostic (CAD) systems has been compared for classification of breast tissue density. The work has been carried out on MIAS dataset with 322 mammographic images consisting of 106 fatty and 216 dense images. The ROIs have been selected from densest region (i.e., the center of each image, ignoring the pectoral muscle) of each mammogram. The total dataset consisted of 322 ROIs (106 fatty ROIs and 216 dense ROIs). Five statistical texture features namely, mean, standard deviation, entropy, kurtosis and skewness are evaluated from Laws' texture energy images resulting from Laws' masks of length 5, 7 and 9. The texture feature vectors computed from Laws' masks of different lengths are then subjected to principal component analysis (PCA) for reduction in feature space dimensionality. The SVM and PNN classifiers are used for the classification task. It is observed that the highest classification accuracy of 92.5 % is achieved with first four principal components derived from texture features computed with Laws' masks of length 7 by using PNN classifier and the highest classification accuracy of 94.4 % is achieved with first four principal components derived from texture features computed with Laws' masks of length 5 by using SVM classifier. It can be concluded that the first four principal components derived from Laws' texture energy images resulting from Laws' masks of length 5 are sufficient to account for textural changes exhibited by fatty and dense mammograms. The promising results obtained by the proposed CAD design indicate its usefulness to assist radiologists for breast density classification.