THE USE OF TRANSFER LEARNING TECHNIQUE IN DIAGNOSING MAMMOGRAM MASSES BASED ON BREAST TISSUE DENSITY

Abstract

Breast cancer is one of the most prevalent cancers, and currently many computers aided detection/diagnosis (CAD) systems are being used in clinical use. Whilst recent studies have shown that there is a high positive correlation between high breast density and high breast cancer risk. Thus, breast density classification may aid in breast lesion analysis. With this objective, we proposed a framework of two systems; the first one classifies the mammographic images into four categories of breast densities. Different sets of features (First order gray-level parameters, Gray-Level co-occurrence matrices, Laws' texture energy measurements and Zernike moment features) were investigated along with several classifiers. The results achieved a promising classification accuracy of 93.7%. While the second system classifies lesions using “Transfer learning” concept based-on pre-trained Convolutional Neural Networks, through investigating and comparing different hyper-parameters to fine-tune several pre-trained models, to find the optimal model configuration proper for each density category.