Radiation dose reduction is one of the most important topics in the field of computed tomography (CT). Over past years, deep learning based denoising methods have been demonstrated effective in reducing radiation dose and improving the image quality. Since the paired low-dose and normal-dose CT are usually not available in clinical scenarios, various learning paradigms are studied, including the fully-supervised learning based on simulation data, the weakly-supervised learning based on unpaired noise-clean or paired noise-noise data, and the self-supervised learning based on noisy data only. Under neither clean nor noisy reference data, unsupervised/self-supervised low-dose CT (LDCT) denoising methods are promising to processing real data and/or images. In this study, we propose the first-of-its-kind Self-Supervised Dual-Domain Network (SSDDNet) for LDCT denoising. SSDDNet consists of three modules including a projection-domain network, a reconstruction layer, and an image-domain network. During training, a projection-domain loss, a reconstruction loss, and an image-domain loss are simultaneously used to optimize the denoising model end-to-end using the single LDCT scan. Our experimental results show that the dual-domain network is effective and superior over single-domain networks in the self-supervised learning setting.
|