Image-Domain CT Metal Artifact Reduction Using an Endpoint-Conditioned Diffusion Bridge
Abstract
Background: Metal implants cause streaking and shading artifacts in computed tomography (CT), obscuring anatomy and distorting Hounsfield-unit (HU) values. Conventional diffusion models reconstruct images from Gaussian noise rather than directly exploiting anatomical information retained in artifact-corrupted CT. Purpose: To develop an image-domain metal artifact reduction (MAR) method connecting artifact-corrupted and metal-free CT images through an endpoint-conditioned diffusion bridge. Methods: The proposed RDBM-MAR method uses an endpoint-conditioned diffusion bridge and initializes reconstruction from the artifact CT image. A conditional U-Net predicts an image-domain correction from intermediate bridge states. Inference uses 10 deterministic reverse steps. We used 14,000 paired CT slices from the AAPM CT-MAR Grand Challenge, including 12,045 training, 633 validation, and 1,322 held-out test slices. One clinical benchmark case was assessed qualitatively. We compared RDBM-MAR with CycleGAN, Pix2Pix, Res-SRDiff, I2I-Mamba, and I2SB using PSNR, SSIM, MAE, and LPIPS. Statistical analysis used Friedman and paired Wilcoxon signed-rank tests with Holm correction. Results: RDBM-MAR achieved the highest PSNR (43.67 +/- 4.56 dB) and SSIM (0.9856 +/- 0.0138) and the lowest MAE (14.74 +/- 8.68 HU). All four metrics differed significantly among methods (p < 0.001). RDBM-MAR outperformed all five comparators in PSNR, SSIM, and MAE (adjusted p < 0.001); Res-SRDiff achieved the lowest LPIPS. Difference maps, HU profiles, and the clinical case showed reduced artifacts with preservation of surrounding anatomy. Conclusions: RDBM-MAR provides a direct restoration trajectory from artifact CT toward metal-free CT, achieving strong structural fidelity and HU accuracy using 10 reverse steps.
Related
- Acquisition-inspired motion artifact modeling and degradation-consistent diffusion restoration in photoacoustic microscopy
- Cone-beam artifact reduction in Gamma Knife CBCT images using a line-arc-line scan trajectory
- Comprehensive reconstruction of collider events with hypergraph representation learning and graph-conditioned diffusion
- Nodal obstruction to conditioned-diffusion representations of the weak momentum
- Diffusion Equation with Riemann–Liouville Derivative in the Angular Domain
BibTeX
@misc{yu2026imagedomain,
title = {{Image-Domain CT Metal Artifact Reduction Using an Endpoint-Conditioned Diffusion Bridge}},
author = {Hang Yu and Mojtaba Safari and Shansong Wang and Mingzhe Hu and Xiaofeng Yang},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.36319},
eprint = {2610.36319},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.36319},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Image-Domain CT Metal Artifact Reduction Using an Endpoint-Conditioned Diffusion Bridge
AU - Hang Yu
AU - Mojtaba Safari
AU - Shansong Wang
AU - Mingzhe Hu
AU - Xiaofeng Yang
PY - 2026
DA - 2026/10/26
PB - arXiv
DO - 10.48550/arxiv.2610.36319
UR - https://arxiv.org/abs/2610.36319
AB - Background: Metal implants cause streaking and shading artifacts in computed tomography (CT), obscuring anatomy and distorting Hounsfield-unit (HU) values. Conventional diffusion models reconstruct images from Gaussian noise rather than directly exploiting anatomical information retained in artifact-corrupted CT. Purpose: To develop an image-domain metal artifact reduction (MAR) method connecting artifact-corrupted and metal-free CT images through an endpoint-conditioned diffusion bridge. Methods: The proposed RDBM-MAR method uses an endpoint-conditioned diffusion bridge and initializes reconstruction from the artifact CT image. A conditional U-Net predicts an image-domain correction from intermediate bridge states. Inference uses 10 deterministic reverse steps. We used 14,000 paired CT slices from the AAPM CT-MAR Grand Challenge, including 12,045 training, 633 validation, and 1,322 held-out test slices. One clinical benchmark case was assessed qualitatively. We compared RDBM-MAR with CycleGAN, Pix2Pix, Res-SRDiff, I2I-Mamba, and I2SB using PSNR, SSIM, MAE, and LPIPS. Statistical analysis used Friedman and paired Wilcoxon signed-rank tests with Holm correction. Results: RDBM-MAR achieved the highest PSNR (43.67 +/- 4.56 dB) and SSIM (0.9856 +/- 0.0138) and the lowest MAE (14.74 +/- 8.68 HU). All four metrics differed significantly among methods (p < 0.001). RDBM-MAR outperformed all five comparators in PSNR, SSIM, and MAE (adjusted p < 0.001); Res-SRDiff achieved the lowest LPIPS. Difference maps, HU profiles, and the clinical case showed reduced artifacts with preservation of surrounding anatomy. Conclusions: RDBM-MAR provides a direct restoration trajectory from artifact CT toward metal-free CT, achieving strong structural fidelity and HU accuracy using 10 reverse steps.
ER -