Image-Domain CT Metal Artifact Reduction Using an Endpoint-Conditioned Diffusion Bridge

‡Corresponding Author: [email protected]

Introduction

Computed tomography (CT) is a cornerstone of clinical imaging, providing detailed visualization of anatomical structures together with quantitative tissue information through the Hounsfield unit (HU) scale. Metal implants, however, violate the assumptions underlying standard reconstruction: beam hardening, photon starvation, scatter, and the resulting inconsistencies in the projection data generate bright and dark streaks that radiate from the metal [1]. These artifacts distort local HU values, obscure the anatomy immediately adjacent to the implant, and propagate into downstream tasks such as segmentation, anatomical localization, and quantitative image analysis. Metal artifact reduction (MAR) has therefore been studied extensively, yet suppressing artifacts while simultaneously preserving local anatomical structure and quantitative CT intensity remains an open problem [1, 2].

Approaches to MAR fall broadly into projection-domain and image-domain families. Projection-domain methods, including linear interpolation [3], normalized MAR [4], and their learning-based successors [5], correct the inconsistent measurements at their physical origin and can therefore be highly effective, but they require access to raw projection data together with the scanner geometry, bowtie filter, and calibration parameters used for reconstruction. Such data are proprietary and are rarely exported in clinical practice, where the archived product is the reconstructed image; models developed for one scanner geometry additionally transfer poorly to another. Errors introduced during sinogram completion also propagate along entire ray paths and can manifest as secondary artifacts remote from the implant [6], whereas corrections applied in the image domain remain spatially local. We therefore adopt an image-domain formulation, which is directly applicable to retrospective and clinical data for which only reconstructed images are available.

Within the image domain, a network learns the mapping from artifact-corrupted CT images to their metal-free counterparts from paired training data. Conditional image-to-image translation models [7] achieve strong performance on paired restoration tasks and have been applied directly to MAR [8], suppressing prominent streaks and restoring the appearance of corrupted regions. Their adversarial objectives, however, are not intensity-preserving by construction, and reported failure modes include local HU shifts, residual streaking, and over-smoothing of fine anatomical detail [8, 6]. This limitation reflects a broader tension between perceptual quality and distortion fidelity [9], and it is consequential in MAR, where the corrected image is frequently used for quantitative measurement.

Diffusion models address part of this problem by modeling restoration as a gradual stochastic transformation between image distributions [10, 11], but in the conventional formulation sampling begins from Gaussian noise and the degraded observation enters only as conditioning information, even though that observation already carries most of the anatomical and intensity content of the target. A more efficient alternative is to let the residual between the paired images drive the process. We previously showed that shifting the forward process along the residual between low- and high-resolution MRI aligns the two distributions directly and permits high-fidelity reconstruction in as few as four sampling steps, at a fraction of the cost of conditional diffusion [12]. Related formulations construct an explicit stochastic bridge between the paired endpoints [13, 14], and allow the residual, where available, to modulate where noise is injected and removed, so that perturbation concentrates on degraded regions while intact anatomy is left largely undisturbed [14].

Metal artifacts differ from the natural-image degradations considered in previous diffusion bridge studies. They are signed and non-local, correlated along ray paths rather than confined to a compact region, and the restoration target is a quantitative HU value rather than a plausible visual appearance. For CT metal artifact reduction, we therefore use the uniform-modulation Ornstein–Uhlenbeck bridge formulation. The paired clean-to-artifact difference remains in the bridge mean, while the Gaussian perturbation is spatially uniform. This provides a direct stochastic transition between the paired metal-free and metal-artifact endpoints without residual-dependent noise modulation.

In the proposed method, the metal-free CT image $x_0$ and its paired metal-artifact image $\mu$ are defined as the two endpoints of the diffusion bridge, following the notation of Wang et al. [14], in which $\mu$ denotes the degraded observation rather than the linear attenuation coefficient. During training, intermediate bridge states are sampled between the endpoints, and a conditional U-Net [15] receives the current bridge state, the metal-artifact image, and the timestep, and estimates the corresponding metal-free CT image. At inference, the reverse process is initialized directly from the observed metal-artifact image and recovers the clean image along a shortened sampling trajectory. Because the network predicts an image-domain correction relative to the degraded input rather than synthesizing a complete CT image from Gaussian noise, the restoration proceeds directly from the observed metal-artifact image toward the metal-free target while retaining the anatomy already present in the input.

Our main contributions are as follows:

• We adapt an endpoint-conditioned Ornstein–Uhlenbeck diffusion bridge to image-domain CT metal artifact reduction, in which the artifact-corrupted scan serves

as a fixed endpoint of the stochastic process and the network predicts an image-domain correction relative to the degraded input, preserving the anatomy and HU content already present in the scan. The contribution is the MAR-specific instantiation and systematic evaluation, rather than the bridge formulation itself.

• We use the uniform-modulation bridge formulation, in which the paired clean-to-artifact difference contributes to the bridge mean while the Gaussian perturbation remains spatially uniform, avoiding residual-dependent stochastic modulation while retaining direct conditioning on the paired endpoints.

• We benchmark the proposed method against representative methods from four distinct restoration families including adversarial (Pix2Pix [7], CycleGAN [16]), diffusion-based (Res-SRDiff [12]), bridge-based (I²SB [13]), and state-space (I2I-Mamba [17]), all trained and evaluated under identical data splits and preprocessing on a common held-out CT-MAR test set.

• We evaluate both fidelity and perceptual quality using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [18], mean absolute error (MAE), and learned perceptual image patch similarity (LPIPS) [19], and report slice-level metric distributions with paired statistical comparisons rather than cohort means alone, so that differences between methods are assessed for consistency as well as magnitude.

• We separate artifact suppression from quantitative intensity recovery through localized difference maps and HU line-profile analysis, addressing a distinction that aggregate image-quality metrics do not capture.

• We assess generalization beyond the simulated training distribution on clinical CT data from the AAPM CT-MAR benchmark [20, 21].

Materials and methods

Data Preparation

The proposed model was developed and evaluated on the paired image-domain CT data released by the AAPM CT Metal Artifact Reduction (CT-MAR) grand challenge [20, 22]. The challenge database was constructed with a hybrid simulation framework [21] that inserts virtual metal objects into clinical CT images drawn from DeepLesion [23] and the UCLH Stroke EIT dataset [24], covering head and body anatomy. Artifact-corrupted projections and their reconstructions were generated with the CatSim simulator in the open-source XCIST toolkit [25], which models quantum noise, beam hardening, and scatter. Each sample therefore comprises a metal-artifact CT image and its spatially registered metal-free reference, enabling supervised training in the image domain.

All images were processed using the same preprocessing procedure. CT values were clipped to the range of $[-1000, 2500]$ HU to standardize the intensity range and retain diagnostically relevant tissue information, followed by linear normalization to $[0, 1]$. The identical normalization procedure was applied to both the metal-artifact input images and their corresponding metal-free references. No spatial resizing was performed, and all images were retained at their native $512 \times512$ resolution.

A total of 12,045 paired samples were used for training, while 633 paired samples were allocated for validation. A separate held-out test set containing 1,322 paired samples was reserved exclusively for final model evaluation and was not used during model training or model selection. The pairing between each metal-artifact image and its corresponding metal-free reference was strictly maintained throughout data partitioning and evaluation.

For qualitative assessment, one clinical case from the AAPM CT-MAR scoring set was additionally evaluated [20, 21]. This case contained a metal implant, had no metal-free ground truth, and was not used during training. Metal regions were delineated from the unclipped HU image by thresholding at 3000 HU and are overlaid in red on all displayed images; the resulting mask was used for visualization only and was never supplied to the model during training or inference.

CT Metal Artifact Diffusion Bridge

For each paired CT sample, the metal-free image is denoted by $x_0$ and the corresponding metal-artifact image by $\mu$. The forward process establishes a stochastic transition from the metal-free image to its degraded counterpart, whereas the reverse process reconstructs the metal-free image beginning from the observed metal-artifact CT.

We adopt the generalized Ornstein–Uhlenbeck bridge [26], in which Doob’s $h$-transform conditions a mean-reverting OU process on the fixed terminal state $x_T = \mu$, eliminating the terminal variance and yielding a point-to-point mapping between the paired images. This formulation is the uniform-modulation case of the generalized bridge family described by Wang et al. [14]: the stochastic modulation term is held at unity, so the Gaussian perturbation is spatially uniform rather than residual-weighted. The resulting process is

Vector (PDF)

where $\theta_t$ is the time-dependent drift coefficient, $\bar{\theta}_{t:T} = \int_t^T \theta_s\,\mathrm{d}s$ is the accumulated drift over the remaining interval, $\lambda$ scales the bridge variance, and $\omega_t$ is a standard Wiener process.

The forward trajectory need not be simulated: an intermediate bridge state admits the closed form

with coefficients

Here $\Theta_t$ weights the metal-free-to-artifact difference $x_0-\mu$ and $\Sigma_t$ sets the perturbation magnitude. The construction satisfies both endpoint conditions: at $t=0$, $\Theta_0=1$ and $\Sigma_0=0$, recovering $x_0$; at $t=T$, $\Theta_T=0$ and $\Sigma_T=0$, so the state collapses to $\mu$ with no residual noise. The drift schedule $\theta_t$ follows a cosine profile [11], with $\lambda=10^{-4}$ and $T=100$ discretization steps. A conditional U-Net receives the current bridge state $x_t$, the metal-artifact image $\mu$, and the corresponding time variable. The network predicts an image-domain correction $\Delta_\phi$, which is combined with the metal-artifact image to obtain the metal-free estimate,

This residual parameterization confines the learning problem to the artifact-induced component of the image, so that anatomy already present in $\mu$ is carried through unaltered. The network is trained by minimizing an $\ell_1$ loss between $\hat{x}_0$ and the paired metal-free reference $x_0$, computed on the normalized $[0,1]$ intensity scale over all image pixels.

Inference uses a shortened trajectory $\{\tau_K,\ldots,\tau_0\}$ of $K=10$ steps selected by uniform spacing from the training discretization, and is initialized directly from the observation,

Because $\Sigma_T=0$, the general update below is undefined at the first step, which is therefore evaluated directly from Eq. (2) with the noise term omitted:

Subsequent states follow the deterministic update

No stochastic term is injected during reverse sampling, so reconstruction is deterministic given $\mu$ and reproducible across runs. The estimate at $\tau_0$ is returned as the metal-artifact-reduced image and mapped back to HU by inverting the normalization of § 2.1.

Algorithm 1 Training of the CT metal artifact diffusion bridge

Require: Paired training set $\mathcal{D} = \{(x_0,\mu)\}$; network $\Delta_{\phi}$; bridge coefficients $\Theta_t, \Sigma_t$ from Eq. (3)

1. repeat

2. Sample a minibatch $(x_0,\mu) \sim\mathcal{D}$

3. Sample $t \sim\mathcal{U}\{0,\ldots,T-1\}$

4. Sample $\epsilon\sim\mathcal{N}(0,I)$

5. $x_t \leftarrow\mu+ \Theta_t(x_0-\mu) + \Sigma_t\epsilon$ ▷ Eq. (2)

6. $\hat{x}_0 \leftarrow\mu+ \Delta_{\phi}(x_t,\mu,t)$ ▷ Eq. (4)

7. $\phi\leftarrow\phi- \eta\nabla_{\phi}\mathcal{L}(\hat{x}_0,x_0)$

8. until convergence

9. return $\phi$

Algorithm 2 Reverse sampling for CT metal artifact reduction

Require: Metal-artifact image $\mu$; trained network $\Delta_{\phi}$; step sequence $\tau_K > \cdots> \tau_0$; coefficients $\Theta_{\tau_i}, \Sigma_{\tau_i}$

1. $x_{\tau_K} \leftarrow\mu$

2. for $i = K, K-1,\ldots,1$ do

3. $\hat{x}_0^{(i)} \leftarrow\mu+ \Delta_{\phi}(x_{\tau_i},\mu,\tau_i)$

4. if $i=K$ then

5. $x_{\tau_{i-1}} \leftarrow\mu+ \Theta_{\tau_{i-1}}\left(\hat{x}_0^{(i)}-\mu\right)$ ▷ Eq. (6); $\Sigma_{\tau_K}=0$

6. else

7. Update $x_{\tau_{i-1}}$ by Eq. (7)

8. end if

9. end for

10. return $x_{\tau_0}$ ▷ $\Theta_{\tau_0}=1$, $\Sigma_{\tau_0}=0 \Rightarrow x_{\tau_0}=\hat{x}_0^{(1)}$

Implementation Details

The forward bridge was discretized into $T=100$ timesteps with bridge variance $\lambda=10^{-4}$. At each training iteration, a timestep $t$ was drawn uniformly from $\{0,\ldots,T-1\}$ and Gaussian noise $\epsilon\sim\mathcal{N}(0,I)$ was used to construct the corresponding intermediate bridge state via Eq. (2). Because a single conditional network is shared across all timesteps, this sampling exposes the same weights to the full range of intermediate corruption levels, and the same trained model is applied at every reverse stage.

The network was trained using direct clean-image supervision. Specifically, the final prediction was parameterized as an additive image-domain correction relative to the observed metal-artifact image,

and optimized using an $L_1$ reconstruction objective,

The model was optimized using Adam [27] with a learning rate of $8 \times10^{-5}$ and exponential decay coefficients $(\beta_1, \beta_2) = (0.9, 0.99)$. Training was performed for 120,000 iterations using a batch size of 10. Gradient clipping with a maximum norm of 1.0 was applied during optimization. An exponential moving average of the model parameters was maintained with a decay factor of 0.995 and used for validation and inference.

At inference the schedule was reduced to $K = 10$ reverse steps, selected by uniform spacing from the training discretization. Reconstruction was initialized from the observed metal-artifact image and no stochastic term was injected, so inference is deterministic for a fixed input and parameter set. All slices were processed at their native $512 \times512$ resolution, and outputs were mapped back to HU by inverting the normalization described in § 2.1 before metric computation.

Comparison Methods

The proposed diffusion bridge model was evaluated against five representative image restoration approaches: Pix2Pix [7], CycleGAN [16], Res-SRDiff [12], I2I-Mamba [17], and I²SB [13]. These methods were selected to represent different image-to-image restoration paradigms and to provide comparisons with both conventional and more recent deep-learning approaches.

Pix2Pix provides a supervised conditional generative adversarial network baseline that directly learns a mapping between paired source and target images. CycleGAN represents an unpaired adversarial translation framework and therefore does not require one-to-one correspondence between source and target images during optimization. Res-SRDiff provides a residual-shifting diffusion-based restoration approach. I2I-Mamba uses selective state-space modeling for image-to-image transformation, while I²SB formulates paired image translation as a Schrödinger bridge between the source and target distributions. Together, these models cover adversarial, diffusion-based, state-space-based, and bridge-based image restoration strategies.

For all comparison methods, the metal-artifact CT images were used as the source images and the corresponding metal-free CT images as the restoration targets. The same training, validation, and held-out test partitions were used across all methods, together with identical HU clipping, intensity normalization, and image resolution. For supervised approaches, the original artifact–metal-free pairing was preserved during training. CycleGAN used the same source and target training sets but did not use their paired correspondence, consistent with its original unpaired training formulation.

Each comparison model was adapted only as required to accommodate the singlechannel $512 \times512$ CT data and the artifact-to-metal-free restoration task while retaining its original learning formulation as closely as possible. Model selection was performed using the corresponding validation set, and final quantitative evaluation was conducted exclusively on the same 1,322 held-out test slices. The identical metal-free reference images and evaluation procedure were used for all methods to ensure direct slice-level comparison.

Quantitative and Statistical Analysis

Quantitative reconstruction performance was evaluated using PSNR, SSIM, MAE, and learned LPIPS. Higher PSNR and SSIM values indicate better performance, whereas lower MAE and LPIPS values indicate better performance.

Each metric was calculated independently for every held-out test slice using the corresponding metal-free CT image as the reference. PSNR and SSIM were computed using the same normalized intensity representation for all evaluated methods. For MAE calculation, reconstructed and reference images were mapped back to the original HU scale, and MAE was therefore reported in HU. For LPIPS evaluation, the single-channel CT images were replicated across three channels and mapped to the input range required by the pretrained LPIPS network. The same evaluation pipeline was applied to all reconstruction methods. Quantitative results were summarized as mean $\pm$ standard deviation, and slice-level metric distributions were visualized using boxplots.

Statistical comparisons were performed using nonparametric paired analyses. For each image-quality metric, differences among the six reconstruction methods were first assessed using the Friedman test for repeated measures, with each held-out test slice treated as a matched observation across methods. The metal-artifact input was not included in the Friedman analysis. Kendall’s coefficient of concordance ($W$) was calculated to quantify the effect size of the overall between-model differences.

Following a significant Friedman test, pairwise comparisons between the proposed method and each of the five comparison models were performed using the Wilcoxon signed-rank test on the paired slice-level metric values. Holm correction was applied separately within each image-quality metric to account for multiple comparisons. An adjusted $p < 0.05$ was considered statistically significant. Median paired differences were additionally reported for the closest-performing comparison method to describe the magnitude and direction of the observed differences.

Results

Qualitative Comparison

Figure 2 and Figure 3 show representative artifact reduction results from the head and thoracic regions. The reconstructed images are compared with the corresponding metal-free references, and the difference maps show the remaining HU deviations after artifact reduction compared to the reference.

Figure 2 presents the representative head CT example. The metal-artifact input contains pronounced streak artifacts extending from the metallic regions into the surrounding anatomy. All evaluated methods reduced the dominant artifacts to varying degrees. CycleGAN and Pix2Pix retained larger residual intensity deviations, whereas Res-SRDiff and I2I-Mamba provided stronger artifact suppression. $I^{2}\mathrm{SB}$ produced a reconstruction closely matching the metal-free reference. The pro posed diffusion bridge model further reduced residual intensity errors while preserving local anatomical structures. The corresponding difference map showed limited HU deviations in the artifact-affected regions.

Figure 3 presents the representative thoracic CT example. The metal-artifact input contains streak artifacts extending from the metallic regions across the surrounding anatomy. CycleGAN and Pix2Pix reduced the major streak patterns but retained visible local intensity differences. Res-SRDiff, I2I-Mamba, and $I^{2}$SB showed stronger artifact reduction and closer agreement with the metal-free reference. The proposed model effectively reduced the streak artifacts while preserving local anatomical structures and soft-tissue contrast. Its difference map also showed reduced residual HU deviations in the artifact-affected regions.

Quantitative and Statistical Comparison

Table 1 summarizes the quantitative performance of the evaluated methods on the 1,322 held-out test slices. The original metal-artifact images had a PSNR of $29.48 \pm3.59$ dB, an SSIM of $0.7972 \pm0.1232$, an MAE of $57.33 \pm22.14$ HU, and an LPIPS of $0.2912 \pm0.1401$.

Table 1: Quantitative comparison on the held-out CT-MAR test set. Values are reported as mean $\pm$ standard deviation. SSIM and LPIPS are displayed in units of $10^{-2}$. Bold values indicate the best mean performance for each metric. Friedman tests compare the six reconstruction methods, excluding the artifact input; ${}^{***}p < 0.001$ indicates an overall difference among methods, not pairwise significance. Kendall’s $W$ quantifies the omnibus effect size.

Method	PSNR (dB) $\uparrow$	SSIM ($\times10^{2}$) $\uparrow$	MAE (HU) $\downarrow$	LPIPS ($\times10^{2}$) $\downarrow$
Input (artifact)	29.48 $\pm$ 3.59	79.72 $\pm$ 12.32	57.33 $\pm$ 22.14	29.12 $\pm$ 14.01
CycleGAN	34.32 $\pm$ 3.72	93.02 $\pm$ 4.59	36.32 $\pm$ 15.24	8.84 $\pm$ 5.68
Pix2Pix	38.54 $\pm$ 3.56	96.25 $\pm$ 2.16	24.19 $\pm$ 10.40	6.03 $\pm$ 3.16
Res-SRDiff	39.89 $\pm$ 3.22	95.00 $\pm$ 2.31	23.15 $\pm$ 9.07	1.50 $\pm$ 1.21
I2I-Mamba	40.41 $\pm$ 3.37	97.52 $\pm$ 1.63	20.57 $\pm$ 9.48	5.13 $\pm$ 4.37
I²SB	43.48 $\pm$ 4.35	98.37 $\pm$ 1.44	16.59 $\pm$ 9.87	1.69 $\pm$ 1.86
RDBM-MAR (ours)	43.67 $\pm$ 4.56	98.56 $\pm$ 1.38	14.74 $\pm$ 8.68	4.01 $\pm$ 5.64
Friedman $\chi^{2}(5)$	$5818.91^{***}$	$5650.74^{***}$	$5980.92^{***}$	$4779.70^{***}$
Kendall’s $W$	0.880	0.855	0.905	0.723

Among the reconstruction methods, our proposed model achieved the highest PSNR of $43.67 \pm4.56$ dB, the highest SSIM of $0.9856 \pm0.0138$, and the lowest MAE of $14.74 \pm8.68$ HU. I{sup}2SB provided the closest performance for these three metrics, with a PSNR of $43.48 \pm4.35$ dB, an SSIM of $0.9837 \pm0.0144$, and an MAE of $16.59 \pm9.87$ HU. I2I-Mamba achieved a PSNR of $40.41 \pm3.37$ dB, an SSIM of $0.9752 \pm0.0163$, and an MAE of $20.57 \pm9.48$ HU.

The ranking differed for LPIPS. Res-SRDiff achieved the lowest LPIPS ($0.0150 \pm0.0121$), followed by I{sup}2SB ($0.0169 \pm0.0186$), whereas the proposed model obtained an LPIPS of $0.0401 \pm0.0564$. These results indicate that the proposed model achieved the strongest performance in image fidelity, structural similarity, and HU accuracy, while Res-SRDiff achieved the lowest perceptual feature-space error.

Figure 4 shows the slice-level distributions of PSNR, SSIM, MAE, and LPIPS across the held-out test set. The proposed model maintained favorable PSNR, SSIM, and MAE distributions across the test slices, while substantial overlap with the strongest comparison methods remained for some metrics. The LPIPS distributions followed a different pattern, with Res-SRDiff and I²SB shifted toward lower values. The slice-level variability and overlap among methods motivated the paired statistical analysis rather than relying solely on the mean values reported in Table 1.

Vector (PDF)

The Friedman omnibus test showed significant differences among the six reconstruction methods for all four image-quality metrics. Significant differences were observed for PSNR ($\chi^{2}(5) = 5818.91$, $p < 0.001$), SSIM ($\chi^{2}(5) = 5650.74$, $p < 0.001$), MAE ($\chi^{2}(5) = 5980.92$, $p < 0.001$), and LPIPS ($\chi^{2}(5) = 4779.70$, $p < 0.001$).

Kendall’s $W$ values were 0.880, 0.855, 0.905, and 0.723 for PSNR, SSIM, MAE, and LPIPS, respectively, indicating substantial between-method differences across the paired test slices.

Pairwise Wilcoxon signed-rank tests with Holm correction showed that the proposed model significantly outperformed each of the five comparison methods in PSNR, SSIM, and MAE (all adjusted $p < 0.001$). Compared with I²SB, the closest-performing comparison method for these metrics, the proposed model achieved median paired differences of $+0.556$ dB in PSNR, $+0.0014$ in SSIM, and $-1.557$ HU in MAE. For LPIPS, the proposed model achieved significantly lower values than CycleGAN, Pix2Pix, and I2I-Mamba, but significantly higher values than Res-SRDiff and $I^{2}$SB after Holm correction.

HU Profile Analysis

Figure 5 provides an HU profile comparison for the same head CT example shown in Figure 2. The metal-artifact input showed substantial deviations from the metal-free reference along this profile. Although all reconstruction methods reduced these deviations to varying degrees, differences remained around the regions with the strongest intensity variation. The profile produced by the proposed model more closely followed the reference CT values across the artifact-affected region. This local profile analysis complements the whole-slice MAE by illustrating how the reconstruction behaves across individual anatomical and artifact boundaries.

Clinical CT Evaluation

Figure 6 shows the application of the proposed model to clinical images from the AAPM CT-MAR scoring dataset. Because a metal-free reference image is not available for this clinical case, the evaluation was conducted qualitatively. The input contains visible metal-associated streak artifacts surrounding the implant regions. After reconstruction, the proposed model reduced the prominent streaking and improved the visibility of the surrounding anatomical structures.

Metal regions identified from the original CT image using a threshold of 3000 HU are displayed as red overlays for visualization only. The metal mask was not provided to the model during training or inference. This clinical example provides an additional qualitative evaluation on clinically representative CT data containing real metal artifacts.

Discussion

The proposed diffusion bridge model demonstrated strong performance for image-domain CT metal artifact reduction across the held-out test set. Among the six reconstruction methods, the proposed model achieved the highest PSNR and SSIM and the lowest MAE, indicating closer agreement with the metal-free reference images. in terms of overall image fidelity, structural similarity, and CT intensity accuracy.

The paired statistical analysis further supported these findings. The Friedman tests showed significant differences among the evaluated methods for all four metrics, with Kendall’s $W$ values ranging from 0.723 to 0.905. Pairwise Wilcoxon signed-rank tests with Holm correction showed that the proposed model significantly outperformed all five comparison methods in PSNR, SSIM, and MAE. These results indicate that the observed improvements were consistent across the paired test slices rather than being driven only by differences in average performance.

The qualitative results were consistent with the quantitative evaluation. In both the representative head and thoracic CT examples, the proposed model substantially reduced metal-induced streak artifacts while preserving the surrounding anatomical structures. The corresponding difference maps showed smaller residual HU deviations in artifact-affected regions. The HU line-profile analysis further showed that the reconstructed CT values preserved the overall intensity trend of the metal-free reference with reduced local fluctuations. Together with the lower MAE observed across the full test set, these findings support the quantitative reliability of HU recovery by the proposed model and suggest that artifact-related intensity errors can be reduced with lower risks of altering the underlying anatomical information.

The ranking of the methods differed for LPIPS. Res-SRDiff achieved the lowest LPIPS, followed by $I^{2}$SB, whereas the proposed model achieved the best performance in PSNR, SSIM, and MAE. Pairwise testing similarly showed that the proposed model had lower LPIPS than CycleGAN, Pix2Pix, and I2I-Mamba, but higher LPIPS than Res-SRDiff and $I^{2}$SB. This difference indicates that the similarity of the learned perceptual features and the accuracy of CT intensity do not necessarily rank the reconstruction methods in the same order, consistent with the perception–distortion trade-off in image restoration [9]. In the reduction of CT metal artifacts, this distinction is important because accurate recovery of HU values and anatomical structures is required in addition to visual similarity. The lower MAE of the proposed method therefore provides complementary information to the LPIPS results. Unlike natural-image restoration, CT reconstruction requires preservation not only of visual appearance but also of quantitatively meaningful image intensities. HU values carry crucial physical information that is used in quantitative image interpretation and clinical analysis. In this context, the lower MAE achieved by the proposed model indicates improved recovery of CT intensity values and complements the information provided by perceptual metrics such as LPIPS. The simultaneous improvement in PSNR, SSIM, and MAE therefore suggests that the proposed method reduces artifact-related errors while maintaining both structural and quantitative fidelity to the metal-free reference.

The observed performance may be related to the formulation of the restoration process. Unlike conventional diffusion approaches that initialize reverse reconstruction from an independent Gaussian sample, the diffusion bridge starts directly from the observed metal-artifact CT image. As a result, anatomical information that remains present in the degraded observation is retained at the beginning of reconstruction rather than being regenerated from noise. The conditional U-Net further predicts an image-domain correction relative to the metal-artifact image, focusing the reconstruction on the differences between the degraded observation and the metal-free target. The bridge formulation therefore provides a direct restoration trajectory between the paired artifact and metal-free endpoints. In addition, the shortened deterministic reverse trajectory allows reconstruction to be performed with a limited number of network evaluations. Although the contribution of each component was not separately evaluated in the current study, these design characteristics may contribute to the strong intensity and structural fidelity observed in the results.

The clinical CT example from the AAPM CT-MAR scoring dataset provided an additional qualitative evaluation beyond the paired benchmark data. The model reduced visible streak artifacts surrounding the metal regions and improved the visibility of nearby anatomical structures. However, a metal-free reference was not available for this case, and the clinical result should therefore be interpreted qualitatively rather than as a quantitative validation of performance on real metal artifacts.

Several limitations remain. The quantitative evaluation was performed using paired benchmark data generated through the AAPM CT-MAR simulation framework, and performance on clinical CT images may differ because of variations in scanner configuration, implant composition, anatomy, and artifact severity. The cur rent model also processes each CT slice independently and does not use information from adjacent slices or three-dimensional anatomical continuity. In addition, the clinical evaluation was limited to a single scoring case without a metal-free reference. Future work will therefore include evaluation on a larger range of clinical CT images and implant types, reader-based assessment, and extension of the framework to adjacent-slice or three-dimensional reconstruction. Metal masks and projection-domain measurements could also be incorporated as additional conditioning information. Finally, alternative training objectives may help improve perceptual similarity while preserving the HU accuracy and structural fidelity achieved by the current model.

Conflicts of interest

There are no conflicts of interest declared by the authors.

Acknowledgment

This research is supported in part by the National Institutes of Health under Award Numbers R01HL183223, R01DE033512, and R01CA272991.

This work used the datasets provided by the AAPM CT-MAR Grand Challenge. The training datasets were generated using a hybrid data simulation framework, which incorporates publicly available clinical images from the DeepLesion CT dataset and the UCLH Stroke EIT Dataset, together with metal objects generated using the CatSim simulator in the open-source toolkit XCIST. This work also used clinical CT data from the benchmark provided by the AAPM CT-MAR Grand Challenge.

The dataset and associated benchmark resources are available through the official challenge website: https://www.aapm.org/GrandChallenge/CT-MAR/Default.asp.