Research On MR Pseudo-normal Brain Image Synthesis Method Based On Generative Adversarial Networks

Brain tumor is one of the incurable diseases and the treatment of brain tumor is facing severe challenges.In order to accelerate the diagnosis process of brain tumor,in recent years,brain tumor segmentation methods based on deep learning has developed rapidly and achieved significant results.The role of deep models is to model the relationship between input images and corresponding segmentation labels.Unlike the deep models,doctors often segment brain tumor accurately by comparing the morphological structure of healthy brain images in the actual diagnosis process.Inspired by this,this paper aims to generate high-quality pseudo-healthy brain images corresponding to brain tumor images.Since there are no healthy brain images corresponding to the brain tumor images in the real medical scene,this paper finishes the generation task based on generative adversarial networks,which has achieved significant performance in unsupervised generation task.The generation task has three main problems:first,there is no healthy brain images corresponding to the brain tumor images in the actual medical scene,so supervised learning method is unsuitable;second,due to the high variability of generative adversarial network,the generated images are easy to be out of controlled;third,there is no full reference images corresponding to the generated images,so it is hard to evaluate the quality of the generated images.In view of the above three problems,this paper introduces pseudo-healthy brain image generation model ANT-GAN based on generative adversarial network and the generative quality evaluation methods.Aiming at the first problem,this paper introduces the cyclic generative adversarial training method,which can achieve good generation results without paired samples;for the second problem,this paper introduces a global connection structure and a healthy region consistency loss function to constrain the generator to only change the tumor region,and keep non-tumor,the healthy region unchanged as much as possible;for the third problem,this paper introduces a variety of methods to evaluate the quality of generated images indirectly or directly,including generated images visualization,t-SNE feature distribution,healthy region PSNR and segmentation/classification performance.The method proposed in this paper is evaluated in the brain tumor dataset Brats 18.The qualitative and quantitative results show that ANT-GAN can achieve significant results on the MR pseudo-healthy brain images generation task,and the generated pseudo-healthy brain images can improve the performance of deep learning-based brain tumor segmentation and classification tasks.Moreover,this paper conducted an experiment on the CT liver lesion dataset LiTS.The experimental results show that ANT-GAN can achieve significant results on CT liver pseudo-healthy generation task,indicating that ANT-GAN can be extended to other pseudo-healthy medical images synthesis tasks.