Detail Recovery in Medical Images Denoising

Details in medical images are important for doctors' diagnosis. However, the images processed by current image denoising methods tend to be over-smooth at edges to some extent due to the lack of special attention to details. In order to remove noises and preserve details of medical images, we propose a GAN-based network named SDGAN which focuses on the recovery of details after denoising. The proposed SDGAN contains two subnetworks: one denoises medical images preliminarily, and the other attempts to reconstruct the details missed in the previous subnetwork. Experiments on Wireless Capsule Endoscopic(WCE) images with noises are conducted to evaluate the performance of SDGAN in comparison with other state-of-the-art denoisers. The results show significant gains in terms of quantitative metrics (PSNR and SSIM) and visual effects using SDGAN. SDGAN is able to recover realistic details textures, making it highly attractive for medical image denoising applications.