BREAST IMAGE CLASSIFICATION USING ENSEMBLE DEEP LEARNING ARCHITECTURES

Abstract

Breast cancer is one of the most widespread diseases around the world which mostly leads to the death not only in women, but also men. The rapid development of the technology and the disruption era, have contributed into finding feasible solutions to diagnose this type of cancer in its early stages. These have made possible the acquirement of high-resolution breast tissue images. On the other hand, Computed Aided Detection (CAD) Systems have taken advantage of the availability of the datasets, and have somehow overcame the struggles in diagnosing the breast cancer. Being difficult to analyze because of their shape, size, and different contrast levels, breast tissues classifications cannot be utilized by the manual or traditional image processing techniques. Considering the afore-mentioned reasons, there is room for computerized implementations for the classification of the breast tissues into malignant or benign. Convolutional Neural Networks (CNN) have assisted in the classification of the abnormalities, producing very accurate results. In this study, we examine the ability of such networks in classifying the breast medical images into two categories: benign or malign. This approach, adapts two pre-trained deep learning architectures such as Densenet201 and VGG16 and ensemble them in a stack learning model. Each of the architectures has been trained separately, and then the outputs of them both, have been fed into a stack model of Multilayer perceptron classifier in order to proceed with the classification of the tissues. For this task, we have used a publicly available dataset of breast histology called BreakHis. It contains a total of 7909 medical images from which 5429 malign and 2480 benign. The proposed ensemble model obtained a higher prediction accuracy than the single implemented classifiers. It achieved accuracy of 91.03%, demonstrating a successful utilization of deep learning architectures in classifying breast tissues.