SEMI-SUPERVISED GAN USING SPARSE SWITCHABLE NORMALIZATION FOR BREAST CANCER CLASSIFICATION

An automatic image recognition system for histopathology, such as a deep neural network (DNN), plays a crucial role in speeding up diagnosis and reducing the error rate. The lack of histopathology data is an obstacle for DNN training, and labeled data collection involves considerable human effort and/or time-consuming experiments. In this paper, we propose a semi-supervised generative adversarial network (SGAN) to diagnose breast cancer from the histopathology images. We incorporate a small amount of labeled data with a large amount of the unlabeled data. Our proposed SGAN creates many fake images to compensate for the lack of images in histopathology images, thus outperforming a traditional convolution neural network (CNN). We apply the sparse switchable normalization (SSN) instead of general batch normalization to improve the performance. Experimental results demonstrate that our proposed model on invasive ductal carcinoma (IDC) dataset significantly improves the performance

Keywords:

breast cancer , image classification , semi-supervised learning , generative adversarial networks , sparse switchable normalization , invasive ductal carcinoma