SIPID: A deep learning framework for sinogram interpolation and image denoising in low-dose CT reconstruction

Huizhuo Yuan, Jinzhu Jia, Zhanxing Zhu

Abstract: Low-dose CT plays a significant role in reducing radiation risks to patients. The main challenge is to achieve better image quality while lowering the imaging dose. In this work, we propose a hybrid deep learning approach that combines sinogram interpolation with image denoising, referred to as SIPID. Through alternatively training the sinogram interpolation network and the image denoising network, the proposed SIPID network can achieve more accurate reconstructions, compared with pure image denoising. We empirically achieved a > 2dB improvement on PSNR based on the Residual U-net denoising structure. Furthermore, we highlight that our design of sinogram interpolation network can be a promising component in CT reconstruction, since it can also seamlessly fit to all kinds of image denoising networks.