VoxelSynth3D: Interpretable Volumetric Image-Domain Metal Artifact Reduction with a Paired Synthetic CLINIC-Metal Benchmark

Amritesh Banerjee; Abdul Basit; Renil Renji Joseph; Nouhaila Innan; Muhammad Shafique

cs.CVOct 1, 2026

VoxelSynth3D: Interpretable Volumetric Image-Domain Metal Artifact Reduction with a Paired Synthetic CLINIC-Metal Benchmark

Authors: Amritesh Banerjee, Abdul Basit, Renil Renji Joseph, Nouhaila Innan, Muhammad Shafique

Organizations: University of Massachusetts at Amherst, Amherst, USA · eBRAIN Lab, Division of Engineering, New York University Abu Dhabi (NYUAD), Abu Dhabi, UAE · Indian Institute of Technology Delhi Abu Dhabi, UAE · Center for Quantum and Topological Systems (CQTS), NYUAD Research Institute, NYUAD, Abu Dhabi, UAE

Abstract

Metal artifacts in postoperative musculoskeletal CT obscure bone-implant and adjacent soft-tissue interfaces. Many metal artifact reduction (MAR) methods require unavailable raw projections or learned models that may shift across scanners and implants. We present VoxelSynth3D, a training-free 3D image-domain framework for reconstructed CT. The framework combines support masking, normalized tissue synthesis, deviation gating, and restricted edge refinement. Detected implant voxels are preserved in the output, while correction targets metal-induced artifacts in the surrounding tissue. We also construct Synthetic CLINIC-Metal, a controlled paired synthetic evaluation resource, from no-metal CTPelvic1K volumes with clean targets, metal/artifact masks, fixed seeds, and patient-level splits; 75 unpaired real metal cases receive qualitative/no-reference evaluation only. The operating point was fixed in a near-flat validation basin. With exact-mask oracle localization, all methods share a metal-excluded tissue ROI. On 40 held-out cases, VoxelSynth3D reduced RMSE from 801.48 to 786.18 HU (paired gain 15.30 HU, 95% CI 11.68-19.23), improving every case and exceeding the evaluated 3D Gaussian smoother by 13.58 HU. Clean-edge agreement decreased next to metal but exceeded input beyond 5 mm. Thus, VoxelSynth3D provides case-consistent within-distribution tissue-error reduction with a localized structural tradeoff. Spacing-aware sensitivity retained aggregate broad-region improvement and identified near-metal calibration as a target.

Figures & tables

Fig. 1: Workflow overview and visual intuition for the proposed VoxelSynth3D pipeline. The method operates directly on reconstructed CT volumes without raw projection data, estimates an implant-adaptive support, synthesizes a normalized non-metal tissue prior, applies deviation-gated local blending, and restricts edge-guided refinement to the artifact support. The component ablation path reports how each design element contributes to the final edge-guided configuration. Detected metal is retained in the reported output.

Method	Proj. phys.	CT only	3D vol.	No train	Paired GT	Real eval.
LI/NMAR [ 2 ]	$\checkmark$	$\times$	$\times$	$\checkmark$	$\triangle$	$\triangle$
Pseudo LI/NMAR	$\triangle$	$\checkmark$	$\times$	$\checkmark$	$\checkmark$	$\triangle$
DuDoNet [ 8 ]	$\checkmark$	$\triangle$	$\times$	$\times$	$\triangle$	$\triangle$
InDuDoNet/+ [ 9 , 10 ]	$\checkmark$	$\triangle$	$\times$	$\times$	$\triangle$	$\triangle$
ACDNet family [ 12 ]	$\triangle$	$\triangle$	$\times$	$\times$	$\triangle$	$\triangle$
VoxelSynth3D (ours)	$\times$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\triangle$

TABLE I: Compact positioning of representative MAR methods and their reported evaluation evidence.

Fig. 2: VoxelSynth3D methodology: The corrupted reconstructed CT volume is processed through four deterministic stages: adaptive metal-support estimation, normalized volumetric tissue-prior synthesis, deviation-gated blending, and restricted edge-guided refinement. Together, these steps localize correction to artifact-corrupted regions, reinject supported high-frequency structure, and avoid global smoothing of unaffected anatomy.

Component	Default value
Metal seed	99.5th percentile of voxels above 400 HU, clipped to 1000–1500 HU
Support	$r_{m}=\mathrm{clip}(1.6r_{\rm eff},8,22)$ mm; $r_{e}=\mathrm{clip}(0.8r_{\rm eff},4,10)$ mm
Valid tissue $Q$	finite non-metal soft tissue, $-300<V<500$ HU; fallback: finite non-metal voxels
Prior/gate	normalized Gaussian prior, $\sigma=4$ mm; $\tau=\mathrm{clip}(2.5\sigma_{Q},25,75)$ HU
Weight $W$	2 mm smoothed macro support, deviation gate, 0.70 Sobel-gradient attenuation
Blend/refinement	$V_{b}=0.95P+0.05V$ ; edge percentile 85; guided $\sigma=1$ mm; $\alpha=0.45$

TABLE II: Default deterministic parameters used for the proposed configuration.

Fig. 3: Evaluation design. Paired synthetic CLINIC-Metal supplies clean targets and implant/artifact masks; real CLINIC-Metal has no paired target and supports only qualitative/no-reference analysis. The primary endpoint excludes the inserted metal core.

Method	RMSE (HU) $\downarrow$	MSE red. (%) $\uparrow$	Err.-var. red. $\uparrow$	EPI $\uparrow$	SSIM $\uparrow$	Time (s) $\downarrow$
Uncorrected Input	801.48 [789.40,813.35]	0.00 [0.00,0.00]	0.000 [0.000,0.000]	0.311 [0.290,0.335]	0.239 [0.232,0.247]	1.66 [1.55,1.78]
2D Gaussian	800.14 [787.89,811.90]	0.34 [0.29,0.39]	$-0.010$ [ $-0.013$ , $-0.007$ ]	0.330 [0.308,0.354]	0.239 [0.232,0.246]	14.58 [14.24,14.95]
3D Gaussian	799.75 [787.38,811.92]	0.43 [0.38,0.49]	$-0.008$ [ $-0.011$ , $-0.004$ ]	0.333 [0.310,0.356]	0.239 [0.232,0.246]	16.52 [16.21,16.87]
Pseudo LI-MAR	801.46 [789.33,813.52]	0.00 [0.00,0.01]	0.000 [0.000,0.000]	0.310 [0.289,0.333]	0.240 [0.232,0.247]	97.40 [79.98,116.52]
Pseudo NMAR	801.48 [788.96,813.54]	0.00 [ $-0.01$ ,0.00]	0.000 [ $-0.001$ ,0.000]	0.310 [0.289,0.333]	0.240 [0.232,0.247]	96.61 [78.43,115.75]
DICDNet ∗	819.35 [807.03,831.40]	$-4.52$ [ $-4.60$ , $-4.44$ ]	0.029 [0.027,0.031]	0.301 [0.278,0.325]	0.273 [0.264,0.282]	65.86 [55.42,78.50]

TABLE III: Held-out synthetic results (

n=40

). Image metrics share the metal-excluded tissue ROI and are mean [95% bootstrap CI]

Fig. 4: Tissue-error and structure analysis. (A) Case-level RMSE reduction relative to input. (B) Tissue RMSE–clean-edge F1 tradeoff at 2–5 mm; marker size denotes runtime. (C–D) Distance-resolved RMSE and clean-target edge F1. IDN ∗ denotes public-checkpoint transfer.

Fig. 5: Synthetic supporting panels. (A) Component ablation path: full-ROI RMSE gain over input with EPI overlaid. (B) Full-ROI severity by artifact size.

Mask	Dice	Tissue RMSE	Outside $\|\Delta\|$	Outside $\Delta$ RMSE
		(HU)	(HU)	(HU)
Erode 2 mm	0.394	792.14	0.09	+2.92
Erode 1 mm	0.749	788.45	0.15	+4.55
Exact-mask oracle	1.000	786.18	0.15	+4.62
Dilate 1 mm	0.814	784.13	0.14	+4.22
Dilate 2 mm	0.609	780.80	0.18	+4.91

TABLE IV: Mask sensitivity and collateral modification. Outside

\Delta

RMSE is prediction minus input RMSE in body tissue outside the fixed artifact ROI.

Fig. 6: Paired synthetic qualitative examples. Clean targets and metal-excluded absolute-error maps (0–1200 HU) are shown; darker colors indicate lower error, and cyan outlines the implant. Cases were selected by held-out gain over the shown comparators.

Fig. 7: Real CLINIC-Metal qualitative comparison (window $-900$ to 1300 HU). Paired clean targets are unavailable, so these outputs provide qualitative/no-reference evidence only.

Explore similar work

GraphMAR: Geometry-Aware Graph Learning Framework for Spatially Adaptive CT Metal Artifact Reduction
May 17, 2026Zilong Li, Chenglong Ma, Yiming Lei +7Cone-Beam Computed Tomography Compression Artifacts
Shanghai Key Lab of Intelligent Information Processing, College of Computer Science and Artificial Intelligence, Fudan University, Shanghai 200433, China · Institute of Science and Technology for Brain-inspired Intelligence, Fudan University, Shanghai 200433, China · College of Computer Science and Technology, Qingdao University, Qingdao 266071, China +2
SCMA: Structure-Conditioned and Metal-Aware Flow Matching for CT Metal Artifact Reduction
Jul 30, 2026Heran Wang, Jianing Sun, Xu Jiang +3Cone-Beam Computed Tomography Compression Artifacts
School of Mathematical Sciences, Capital Normal University, Beijing 100048, China · College of Mathematics Science, Inner Mongolia Normal University, Hohhot 010022, China · National Center for Applied Mathematics Beijing, Capital Normal University, and the Academy for Multidisciplinary Studies, Capital Normal University, Beijing 100048, China
Splat-based Metal Artifact Reduction in Cone-Beam CT via Polychromatic Modeling
Aug 13, 2026Kiseok Choi, Inchul Kim, Jaemin Cho +2Cone-Beam Computed Tomography Compression Artifacts
KAIST

May 17, 2026, cs.CVJ/K move · Enter open · S save

GraphMAR: Geometry-Aware Graph Learning Framework for Spatially Adaptive CT Metal Artifact Reduction

Zilong Li, Chenglong Ma, Yiming Lei, Yuanlin Li+6

Shanghai Key Lab of Intelligent Information Processing, College of Computer Science and Artificial Intelligence, Fudan University, Shanghai 200433, China · Institute of Science and Technology for Brain-inspired Intelligence, Fudan University, Shanghai 200433, China · College of Computer Science and Technology, Qingdao University, Qingdao 266071, China+2

Cone-Beam Computed Tomography Compression Artifacts Geometric Reconstruction Stable Anthropometric Priors

Open paper AskPDF arXiv

Computed tomography (CT) metal artifact reduction (MAR) aims to reduce the severe streaking artifacts induced by metallic implants and other high-density objects. Effective MAR generally requires both accurate artifact localization and artifact removal. Sinogram-domain methods can exploit explicit geometric cues, such as metal traces, to identify metal-corrupted measurements, while requiring raw projection data, which is often unavailable in clinical and practical scenarios. Image-domain methods are more flexible and widely applicable, yet they usually lack comparable geometric guidance, limiting their ability to localize artifacts and leading to suboptimal results. To address this limitation, we propose GraphMAR, a geometry-aware learning framework for explicit artifact identification and spatially adaptive MAR in the image domain. The key idea is to introduce graph-based geometric modeling as an image-domain analogue of sinogram metal traces. Specifically, we first construct a geometric graph from the metal mask and derive a geometric density graph that coarsely localizes artifact-prone regions according to inter-implant geometry. We then design GraphMoE, a graph-routed mixture-of-experts module that builds a polar-coordinate artifact graph in feature space and adaptively routes different experts to different spatial regions for MAR. By aligning the learned routing maps with the geometric density graph, GraphMAR provides explicit and interpretable artifact localization while enabling region-adaptive artifact reduction. Experiments on both simulated and real-world datasets demonstrate that GraphMAR achieves superior MAR performance compared with existing methods. To the best of our knowledge, this is the first work to introduce graph-based modeling for CT MAR and to enable explicit artifact identification in the image domain, improving both restoration quality and interpretability.

Abstract

Explore similar work

GraphMAR: Geometry-Aware Graph Learning Framework for Spatially Adaptive CT Metal Artifact Reduction

SCMA: Structure-Conditioned and Metal-Aware Flow Matching for CT Metal Artifact Reduction

Splat-based Metal Artifact Reduction in Cone-Beam CT via Polychromatic Modeling