Medical Image Augmentation Using Image Synthesis with Contextual Function

Xu Yin, Yan Li, Xu Zhang, Byeong-Seok Shin

Abstract: Deep learning technology has been widely used in medical research. For medical images that normally contain more complicated distributions than ordinary images, existing methods have tended to show poor generality when dealing with images of diverse distributions. In recent years, the new method of generative model has begun to receive more and more attention. In this paper, we focus on applications of generative models in medical imaging. We propose a framework with a new contextual loss function that can preserve contexts better than traditional methods. Then we treat it as a data augmentation operation and successfully apply this framework to medical image segmentation. Experiments with generated images and segmentation show that our method is accurate and robust for maintaining semantics, outperforming two existing models under comparison.