Synthesis of brain tumor MR images for learning data augmentation

Abstract

Medical image analysis using a deep learning has been actively researched. The deep learning uses multiple layers from learning data to extract feature maps. Therefore, learning data is an important element of the deep learning. The learning data must have sufficient amount, good quality and wide variety. However, in medical images, it is difficult to secure sufficient amount of data because of patient recruitment, annotation of lesion by experts, and invasion of patient's privacy. To overcome these problems, classic data augmentation methods such as rotation, flip, jittering in brightness and contrast have been used. However, classical augmentation methods have limitations in ensuring data diversity. In order to ensure the diversity of data, there are several generative network-based methods which can generate new data, especially brain tumor MR images. However generative network-based methods are unstable in network training and generated data are generally blurry. It can generate new brain tumor MR images, but they may not be available for supervised learning due to the absence of tumor masks in augmentation of brain tumor images. And in case of augmentation with tumor masks together, there is a limitation of the variety of data since the new tumor masks are generated using simple affinement from the learning data's tumor masks. We propose a method to generate diverse brain tumor MR images that can be used for tumor segmentation. Through simplification of tumor's complex characteristics into a concentric circle, various tumor masks can be generated by controlling the concentric circles which are intuitive and user-friendly objects. By synthesizing the tumor images with the normal brain images, our method generates realistic brain tumor images than the classical augmentation methods and other generation network-based augmentation methods. A comparison study in data augmentation demonstrated that the quality of the images generated from the proposed model outperforms that from the state-of-the-art methods.