[Paper Review] Dual-domain Attention-based Deep Network for Sparse-view CT Artifact Reduction

Xiang Gao, Ting Su+8 more|arXiv (Cornell University)|Mar 17, 2022

Medical Imaging Techniques and Applications4 citations

TL;DR

This paper proposes DDANet, an end-to-end dual-domain attention-based deep learning network that jointly processes CT images and sinograms to reduce streaking artifacts in sparse-view CT reconstruction. By fusing features from both domains through a specialized attention module, DDANet achieves superior artifact suppression and structural preservation, outperforming state-of-the-art methods in numerical, phantom, and in vivo experiments without fine-tuning on real data.

ABSTRACT

Due to the wide applications of X-ray computed tomography (CT) in medical imaging activities, radiation exposure has become a major concern for public health. Sparse-view CT is a promising approach to reduce the radiation dose by down-sampling the total number of acquired projections. However, the CT images reconstructed by this sparse-view imaging approach suffer from severe streaking artifacts and structural information loss. In this work, an end-to-end dual-domain attention-based deep network (DDANet) is proposed to solve such an ill-posed CT image reconstruction problem. The image-domain CT image and the projection-domain sinogram are put into the two parallel sub-networks of the DDANet to independently extract the distinct high-level feature maps. In addition, a specified attention module is introduced to fuse the aforementioned dual-domain feature maps to allow complementary optimizations of removing the streaking artifacts and mitigating the loss of structure. Numerical simulations, anthropomorphic thorax phantom and in vivo pre-clinical experiments are conducted to verify the sparse-view CT imaging performance of the DDANet. Results demonstrate that this newly developed approach is able to robustly remove the streaking artifacts while maintaining the fine structures. As a result, the DDANet provides a promising solution in achieving high quality sparse-view CT imaging.

Motivation & Objective

• To address the challenge of severe streaking artifacts and structural loss in sparse-view CT, which arises from reduced projection data and limits low-dose imaging applications.
• To overcome the limitations of single-domain deep learning methods that process only image or sinogram data in isolation.
• To develop a dual-domain network that leverages complementary information from both image and projection domains for improved reconstruction quality.
• To design an attention mechanism that effectively fuses multi-level features from both domains to enhance artifact removal and structural fidelity.
• To validate the method across diverse datasets—numerical simulations, anthropomorphic phantoms, and in vivo preclinical experiments—without fine-tuning on real data.

Proposed method

• A dual-branch U-Net architecture is used to extract high-level features from the CT image domain and the sinogram domain in parallel.
• The sinogram domain branch employs a fully connected layer followed by an atrous spatial pyramid pooling (ASPP) module to capture multi-receptive field contextual information.
• A novel dual-domain attention module fuses features from both branches, enabling complementary optimization of artifact reduction and structural preservation.
• Dilated convolutions are applied to expand the receptive field and enhance feature representation in both sub-networks.
• The network is trained end-to-end using a loss function that combines L1 and perceptual losses to preserve structural details and reduce artifacts.
• Domain transformation between sinogram and image spaces is facilitated via a fully connected layer to align features across domains.

Experimental results

Research questions

• RQ1 Can a dual-domain deep learning framework that jointly processes CT images and sinograms outperform single-domain methods in reducing streaking artifacts in sparse-view CT?
• RQ2 How effective is an attention-based feature fusion mechanism in combining complementary information from image and projection domains for improved reconstruction?
• RQ3 Does the proposed DDANet maintain fine anatomical structures while suppressing artifacts in low-dose, sparse-view CT imaging?
• RQ4 How generalizable is the DDANet model trained on numerical simulations when applied directly to real phantom and in vivo experimental data?
• RQ5 Can the dual-domain attention mechanism improve image resolution and modulation transfer function (MTF) compared to conventional iterative and deep learning methods?

Key findings

• DDANet significantly reduces streaking artifacts in sparse-view CT reconstructions across all test modalities—numerical simulations, anthropomorphic thorax phantoms, and in vivo monkey head scans—outperforming FBP, TV, and other deep learning baselines.
• In visual comparisons, DDANet preserves fine structural details better than methods like TV and FrameUnet, which produce blurrier images despite reducing artifacts.
• Quantitative MTF analysis confirms that DDANet maintains higher spatial resolution than FBP and TV, with only a slight degradation compared to full-view FBP, indicating strong preservation of high-frequency details.
• The network achieves robust performance on real experimental data without fine-tuning, demonstrating strong generalization from simulated training data.
• Ablation studies confirm the effectiveness of the dual-domain attention module and the contribution of each component, including ASPP and dilated convolutions, to overall performance.
• The method shows consistent superiority in both visual quality and quantitative metrics (e.g., SSIM, PSNR, RMSE) across all evaluation settings.