CT Super Resolution Using Generative Adversarial Network

Deep-learning-based super-resolution (SR) approaches have been emerged to enhance image resolution in CT images, which facilitate accurate disease diagnosis with low dose CT. Although a variety of deep neural networks have been developed, most of them put emphasis on the PSNR of image quality rather than perception quality. Motivated by the super-resolution generative adversarial network (SRGAN), we reconstruct CT images using two proposed advanced CT SR models based on GAN named single-scale CT SRGAN(SCTSRGAN) to achieve pleasing reconstructed results. In the case of single-scale SR, we propose SCT-SRGAN for more powerful feature expression. Specifically, backward fusion refinement module (BFRM) is developed to capture and fuse multi-scale informative feature with dual attention mechanism. Then, we present a new upsampling layer to further enhance the details of reconstructed results. Furthermore, we design a noise reduction block named Noise Reduction Network (NRN) to enrich the textures as well as remove noise and artifacts. Extensive experiments on the public CT image dataset demonstrate that our models achieve favorable performance against state-of-the-art approaches in terms of both objective and subjective metrics efficiently.