A Deep Convolutional Generative Adversarial Network-Based Model to Analyze Histopathological Breast Cancer Images

Abstract

Breast cancer is one of the severe cancers, and early detection is needed to remedy the severity. With the advancement of AI technology, machine learning and deep learning are performing a very important role by automatically finding the types of tumors. A machine learning model needs many data to learn the features of the dataset more precisely. Data scarcity can lead to weak accuracy and can introduce bias toward the higher class. In this study, the BreakHis dataset has been used which contains histopathological images of breast tissues of different magnification factors. The datasets have two classes, but the distribution of images is quite imbalanced. The dataset is balanced by adding generated images to the minority class by using a deep convolutional generative adversarial network (DCGAN). For the classification, four pre-trained deep convolutional neural networks (deep CNN), namely DenseNet, MobileNet, ResNetV2 and Xception, have been applied. After applying DCGAN, the performance has been improved with a maximum increase of 2.92%. As for the classification, the DenseNet model with DCGAN has given the overall top performance in this study. Moreover, the result of this study has outperformed most of the state-of-the-art works in identifying malignancy of breast tissue from histopathological images.