Two-stage Generative Adversarial Recovery Network for MR Brain Images Containing Tumors

Brain image registration (BIR) plays an important role in neuroscience. However, for the registration of brain image containing tumors, the existence of tumor could cause great influence to BIR. One possible solution for getting rid of such influence is to recover the tumor brain image to “normal” appearance brain image (no tumor). Most of existing methods for tumor brain image recovery are based on low-rank, which is time consuming and low recovery quality. In this paper, we propose a novel deep-learning based method for tumor brain image recovery. Specifically, a two-stage generative adversarial network comprising a region recovery stage and an image recovery stage is presented. For the input tumor brain image, the region recovery stage first generates a recovered brain region image containing three different regions (i.e., the gray matter, the white matter and the cerebrospinal fluid). The recovered brain region image is used in the image recovery stage as priori information to get the final “normal” appearance brain image. Both stages are trained under the generative adversarial framework. The experimental results demonstrate that the registration accuracy of tumor brain images can be significantly enhanced by our network as compared to the state-of-the-art image recovery methods.