[Paper Review] DuDoTrans: Dual-Domain Transformer Provides More Attention for Sinogram Restoration in Sparse-View CT Reconstruction

Ce Wang, Kun Shang+4 more|arXiv (Cornell University)|Nov 21, 2021

Medical Imaging Techniques and Applications43 references20 citations

TL;DR

DuDoTrans proposes a dual-domain Transformer architecture that enhances sparse-view CT reconstruction by jointly restoring sinograms using a Sinogram Restoration Transformer (SRT) module and reconstructing CT images via a Residual Image Reconstruction Module. By leveraging long-range dependencies through Swin Transformer-based attention and enforcing consistency between domains, DuDoTrans achieves state-of-the-art performance with fewer parameters and superior robustness across noise levels and datasets.

ABSTRACT

While Computed Tomography (CT) reconstruction from X-ray sinograms is necessary for clinical diagnosis, iodine radiation in the imaging process induces irreversible injury, thereby driving researchers to study sparse-view CT reconstruction, that is, recovering a high-quality CT image from a sparse set of sinogram views. Iterative models are proposed to alleviate the appeared artifacts in sparse-view CT images, but the computation cost is too expensive. Then deep-learning-based methods have gained prevalence due to the excellent performances and lower computation. However, these methods ignore the mismatch between the CNN's extbf{local} feature extraction capability and the sinogram's extbf{global} characteristics. To overcome the problem, we propose extbf{Du}al- extbf{Do}main extbf{Trans}former ( extbf{DuDoTrans}) to simultaneously restore informative sinograms via the long-range dependency modeling capability of Transformer and reconstruct CT image with both the enhanced and raw sinograms. With such a novel design, reconstruction performance on the NIH-AAPM dataset and COVID-19 dataset experimentally confirms the effectiveness and generalizability of DuDoTrans with fewer involved parameters. Extensive experiments also demonstrate its robustness with different noise-level scenarios for sparse-view CT reconstruction. The code and models are publicly available at https://github.com/DuDoTrans/CODE

Motivation & Objective

• Address the mismatch between CNNs’ local feature extraction and sinograms’ global sampling characteristics in sparse-view CT reconstruction.
• Overcome artifacts and noise in low-dose CT by restoring informative sinograms before image reconstruction.
• Improve reconstruction quality and robustness by integrating both sinogram and image domain optimization in a unified Transformer-based framework.
• Ensure consistency between restored sinograms and reconstructed images through a differentiable consistency layer.
• Demonstrate generalizability and computational efficiency across diverse datasets and noise conditions.

Proposed method

• Propose a Sinogram Restoration Transformer (SRT) module based on Swin Transformer’s shifted window self-attention to model long-range dependencies in sinograms.
• Integrate SRT with a Residual Image Reconstruction Module (RIRM) to jointly optimize image reconstruction using both raw and enhanced sinograms.
• Design a differentiable Dual-Domain Consistency Layer to align reconstructed images with restored sinograms, reducing drift error.
• Combine SRT and RIRM into a unified framework, DuDoTrans, enabling end-to-end training with dual-domain supervision.
• Adaptively leverage CNN and Transformer strengths by using patch-based operations with local inductive bias and global context modeling.
• Train and evaluate DuDoTrans on NIH-AAPM and COVID-19 datasets, comparing against CNN-based, Transformer-based, and deep-unrolling methods.

Experimental results

Research questions

• RQ1 Can a Transformer-based sinogram restoration module improve sparse-view CT reconstruction by capturing long-range dependencies in sinogram data?
• RQ2 How does integrating sinogram restoration with image reconstruction in a dual-domain framework affect reconstruction quality and robustness?
• RQ3 Does the proposed SRT module generalize across different reconstruction architectures, including CNNs, Transformers, and deep-unrolling methods?
• RQ4 Can the differentiable consistency layer effectively reduce domain drift and improve alignment between restored sinograms and reconstructed images?
• RQ5 How does DuDoTrans compare in performance, parameter efficiency, and inference speed to state-of-the-art methods under varying noise and view-sparse conditions?

Key findings

• DuDoTrans achieves a PSNR of 32.68 dB on the NIH-AAPM dataset at α_max = 96, outperforming ImgTrans (32.50 dB) and FBPConvNet (31.47 dB).
• On the same dataset, DuDoTrans improves SSIM to 0.9047, surpassing ImgTrans (0.9010) and PDNet+SRT (0.9045).
• With Poisson noise levels of 1e5, 5e5, and 1e6, DuDoTrans maintains the best performance across all α_max settings, showing robustness to noise.
• DuDoTrans-L1 and L2 achieve 0.8–1.0 dB improvement over baseline models with fewer FLOPs, demonstrating computational efficiency.
• On the out-of-distribution COVID-19 dataset, DuDoTrans improves PSNR by 4–5 dB over CNN-based methods and exceeds ImgTrans by 0.4 dB.
• Ablation studies confirm that adding the SRT module improves PSNR by 0.5–1.0 dB across CNN, Transformer, and deep-unrolling baselines.