DDF: A Dual-Domain Fusion Network for Sparse-View CT Reconstruction

Abstract

: Sparse-view computed tomography (SVCT) is a critical technique for reducing radiation exposure, yet it inherently suffers from severe reconstruction artifacts. While dual-domain deep learning methods have advanced SVCT restoration, prevailing cascaded and parallel architectures are limited by one-way information flow or shallow interaction between the two domains. To address this limitation, this work proposes a Dual-Domain Fusion (DDF) framework that establishes a physics-guided, closed-loop feedback mechanism. Unlike conventional pipelines, DDF explicitly maps the intermediate image-domain prediction back to the projection domain via forward projection. Through tailored Sine Fusion and computed tomography (CT) Fusion modules, DDF adaptively integrates structurally rich image feedback with features extracted from the measured sinogram, enabling complementary information exchange and mitigating sinogram information loss. The framework is instantiated using ViewTrans and the Nonlinear Activation Free Network (NAFNet) as the projection-and image-domain backbones, respectively, within a modular formulation that facilitates flexible backbone substitution. Experiments on the 2016 Low-Dose CT Grand Challenge dataset demonstrate strong performance under severe sparsity. At a sparsity factor of 12, DDF achieves a peak signal-to-noise ratio (PSNR) of 36.71 dB and outperforms the matched cascaded architecture by 5.91 dB on the fixed 200-slice test set.