Back to all papers

DuF-DiT: Dual-Domain Fusion Diffusion Transformer with Latent Conditional Modeling for CT Metal Artifact Reduction.

November 19, 2026pubmed logopapers

Authors

Gong Y,Yang S,You M,Tang W,Guo J

Affiliations (5)

  • Sichuan University College of Computer Science, No. 24, South Section 1, Yihuan Road, Chengdu, 610065, China.
  • College of Computer Science, Sichuan University, No. 24, South Section 1, Yihuan Road, Chengdu, 610065, China.
  • State Key Laboratory of Oral Diseases, National Center for Stomatology, National Clinical Research Center for Oral Diseases, Department of Oral Medical Imaging, Sichuan University West China Hospital of Stomatology, No. 14, Section 3, Renmin South Road, Chengdu, Sichuan, 610041, China.
  • Department of Oral and Maxillofacial Surgery, Sichuan University West China Hospital of Stomatology, No. 14, Section 3, Renmin South Road, Chengdu, Sichuan, 610041, China.
  • School of Artificial Intelligence, Sichuan University, No. 24, South Section 1, Yihuan Road, Chengdu, 610065, China.

Abstract

Metal implants introduce nonlinear corruptions into computed tomography (CT) projection measurements, producing severe artifacts in reconstructed images. Dual-domain metal artifact reduction (MAR) methods combine sinogram and image information to correct corrupted projections and suppress image-domain artifacts. However, their cross-domain communication commonly relies on predefined projection operations or fixed interaction pathways. In diffusion-based MAR, sustained interaction between evolving image and sinogram representations throughout the denoising process remains insufficiently explored.
Methods: We introduce the Dual-Domain Fusion Diffusion Transformer (DuF-DiT), which iteratively fuses parallel image and sinogram latent representations during denoising. Its Dual-Domain Fusion Transformer blocks combine intra-domain self-attention with bidirectional cross-attention. A difference-guided gating mechanism adaptively blends conditions from metal-affected observations and linear-interpolation (LI) priors. Denoising in a CT-adaptive latent space further reduces the computational cost of the two parallel streams.
Results: On synthesized DeepLesion data, DuF-DiT achieved the highest mean structural similarity index (SSIM), competitive peak signal-to-noise ratio (PSNR) and root mean square error (RMSE), and the lowest mean absolute error (MAE) in Hounsfield units (HU), with a small mean HU bias. It also achieved the lowest Fréchet Inception Distance (FID) among the compared methods. On clinical CT data, qualitative comparisons showed reduced visible artifacts with preserved anatomical structures, while DuF-DiT yielded the numerically highest mean bone-segmentation Dice coefficient on CLINIC-metal. Ablation studies supported the complementary contributions of adaptive conditioning and bidirectional cross-domain interaction.
Conclusions: The results support sustained dual-domain interaction in latent space as an effective strategy for CT metal artifact reduction, with favorable image fidelity, CT-value accuracy, and downstream bone-segmentation performance under the evaluated conditions.

Topics

Journal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.