PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.01512v1cs.CV

TL;DR

Postoperative musculoskeletal CT artifacts obscure bone–implant and soft-tissue interfaces, while many MAR methods rely on raw projections or training conditions that may not transfer. VoxelSynth3D applies a training-free, localized 3D correction to reconstructed CT, and the paper builds a paired synthetic evaluation resource. On held-out cases it reduces tissue error versus input and an evaluated Gaussian smoother, with a localized edge tradeoff.

  • Problem

    Metal artifacts obscure clinically relevant interfaces in postoperative musculoskeletal CT, and many MAR methods require unavailable raw projections or learned models whose behavior may shift across scanners and implants.

    Source: §I · p. 1

  • Method

    VoxelSynth3D combines implant-adaptive support, normalized tissue synthesis, deviation-gated blending, and restricted edge refinement for reconstructed CT, alongside paired Synthetic CLINIC-Metal evaluation data.

    Source: §I · p. 1

  • Results

    On 40 held-out cases, VoxelSynth3D reduced metal-excluded tissue error versus input and the evaluated 3D Gaussian smoother, with edge loss concentrated near metal.

    Source: §I · p. 2, B. Ablation, Severity, and Metric… · p. 6

  • Takeaways & Limitations

    The supported conclusion is inspectable, training-free reconstructed-volume correction with within-distribution tissue-error reduction and limited robustness near metal.

    Source: §VI · p. 7, F. Limitations · p. 7

    Blinded radiologist scoring remains the key clinical validation step; AAPM and real input-relative metrics cannot establish diagnostic benefit.

    Source: F. Limitations · p. 7

Abstract

from arXiv · show

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.

I. INTRODUCTION

VoxelSynth3D targets metal artifacts in reconstructed musculoskeletal CT with a training-free, localized 3D correction, alongside a paired synthetic evaluation resource. Its reported gains are within-distribution and come with a localized structural tradeoff.

I. INTRODUCTION

  • Postoperative CT artifacts obscure bone–implant and soft-tissue interfaces needed to assess complications, making metal artifact reduction practically important.Beam hardening, scatter, and photon starvation produce streaks and dark bands,.Source: §I · p. 1
  • Projection-domain methods often assume raw projections, while learned MAR can depend on matched training and deployment conditions.Reconstructed-volume CT motivates volumetric, localized methods that do not require unavailable projections or opaque learned behavior.Source: §I · p. 1
  • VoxelSynth3D combines implant-adaptive support, a normalized tissue prior, deviation-gated blending, and restricted edge-guided refinement while preserving detected metal.The correction targets surrounding tissue, and its support, prior, and weights are inspectable.Source: §I · p. 1, Figure 1 · p. 2
  • Synthetic CLINIC-Metal provides paired evaluation from no-metal CTPelvic1K volumes, with clean targets, masks, spacing metadata, and patient-level splits.Real CLINIC-metal scans are kept separate for qualitative and no-reference validation.Source: §I · p. 1
  • On 40 held-out cases, VoxelSynth3D reduced tissue error versus input and the evaluated 3D Gaussian smoother, improving every case while incurring a localized clean-edge cost.The reported claim is within-distribution artifact suppression, not anatomical recovery.Source: §I · p. 2

II. RELATED WORK

Prior MAR approaches offer interpretable projection-domain methods and increasingly capable learned or hybrid alternatives, each with assumptions that matter when only reconstructed CT is available. Synthetic CLINIC-Metal complements broader benchmarks with paired 3D pelvic volumes and spatial masks.

A. Classical and Projection-Domain MAR

  • Linear interpolation and NMAR are interpretable projection-domain methods, but their strongest forms require scanner projection data; reconstructed-image pseudo-sinograms are approximate and warrant caution.LI can erase structures crossing metal traces and introduce streaks.Source: A. Classical and Projection-Domain MAR · p. 2

B. Learning-Based and Hybrid MAR

  • Learned MAR includes dual-domain models such as DuDoNet and InDuDoNet [9],, but performance may depend on matching training and deployment domains.Supervised methods may be sensitive to scanner geometry, implant appearance, and reconstruction kernels.Source: B. Learning-Based and Hybrid MAR · p. 2
  • Reconstructed-image methods remain relevant to retrospective and multi-site studies when raw projections, vendor reconstructions, and kernel details are unavailable.Commercial scanner MAR and orthopedic workflows are clinically useful when available.Source: B. Learning-Based and Hybrid MAR · p. 2
  • Synthetic CLINIC-Metal complements the 2025 AAPM CT-MAR benchmark with paired 3D pelvis volumes and spatial masks for localization analysis.The AAPM benchmark is standardized, 2D, and multi-anatomy –.Source: B. Learning-Based and Hybrid MAR · p. 2

III. METHODOLOGY

VoxelSynth3D applies a deterministic four-stage correction to reconstructed CT, combining adaptive support, tissue-prior synthesis, deviation-gated blending, and restricted edge refinement. The proposed edge-guided configuration is used in the main experiments.

III. METHODOLOGY

  • The proposed edge-guided configuration is the main experimental method; earlier components and later residual or external-cleanup variants are descriptive ablations.The method operates in 3D so correction weights vary smoothly across adjacent axial slices.Source: §III · p. 3
  • The corrupted reconstructed CT passes through adaptive support estimation, normalized tissue-prior synthesis, deviation-gated blending, and restricted edge-guided refinement.Together, these stages localize correction to artifact-corrupted regions, reinject supported high-frequency structure, and avoid global smoothing of unaffected anatomy.Source: §III · pp. 2–3, Figure 2 · p. 3
  • Table II lists the default deterministic parameters used for the proposed configuration.Source: Table II · p. 3, §III · p. 3

A. Adaptive Metal and Artifact Support

Support is defined using implant masks and an implant-size-scaled local region, while a normalized prior estimates tissue intensity from nearby valid non-metal voxels.

A. Adaptive Metal and Artifact Support

  • Synthetic cases use saved inserted-metal masks for exact-mask localization; real CT uses cleaned adaptive HU masks, and implant-size-derived radius scales local correction support.A broader macro support captures dark-band and streak influence, while a narrower edge support limits high-frequency refinement.Source: A. Adaptive Metal and Artifact Support · p. 3

B. Normalized Volumetric Tissue Prior

  • The normalized tissue prior estimates local intensity from valid nearby voxels while excluding metal, nonfinite values, and selected non-soft-tissue extremes.This avoids contamination from extreme metal HU values; a denominator guard and median-tissue fallback address insufficient local support.Source: B. Normalized Volumetric Tissue Prior · p. 3

C. Deviation-Gated Blending

Deviation-gated blending limits replacement to supported, artifact-corrupted regions, and edge-guided refinement partially mitigates local softening without changing distant anatomy.

C. Deviation-Gated Blending

  • A robust threshold estimated from non-metal tissue variation gates blending, while the spatial weight decays with implant distance and is attenuated near strong uncorrupted gradients.Voxels outside the adaptive support remain unchanged.Source: C. Deviation-Gated Blending · p. 3

D. Edge-Guided Refinement

  • Edge-guided refinement uses gradients from the original input and a local linear model to reinject high-frequency structure only at eligible high-gradient voxels within correction support.The conservative blend partially mitigates local softening and does not modify distant anatomy; detected metal voxels remain unchanged.Source: D. Edge-Guided Refinement · pp. 3–4
  • The multi-scale residual-matching extension improves edge metrics but has higher runtime and slightly worse primary RMSE, so it is evaluated only as an ablation.Source: D. Edge-Guided Refinement · p. 4

IV. EXPERIMENTAL SETUP

The spacing-aware sensitivity evaluation retained average broad-region improvement under altered artifact physics but exposed weaker near-metal robustness. Synthetic CLINIC-Metal supplies paired pelvic volumes and masks for controlled within-distribution evaluation.

A. Paired Controlled Synthetic Evaluation Resource

  • Synthetic CLINIC-Metal is a controlled paired resource built from no-metal CLINIC pelvic CT, retaining clean CT as ground truth and enabling reproducible within-distribution benchmarking, not calibrated scanner-physics evaluation.Source-patient splits were assigned before implant insertion: 124 development, 42 validation, and 40 test cases.Source: A. Paired Controlled Synthetic… · p. 4

B. Artifact-Model Sensitivity and Scope

  • Under the spacing-aware artifact model, mean method/input RMSE ratio was 0.891, with improvement in 13/21 cases; the 2–5-mm RMSE ratio was 2.772 and local SSIM changed by −0.150.The analysis identifies the immediate peri-metal region as the principal target for further physics calibration and refinement.Source: B. Artifact-Model Sensitivity and Scope · p. 4

C. Real CLINIC-Metal Validation

Real CLINIC-Metal supports qualitative and no-reference analysis because it lacks paired clean reconstructions. The evaluation compares fixed correction methods using a shared metal-excluded tissue ROI and paired statistical tests.

C. Real CLINIC-Metal Validation

  • The 75 real CLINIC-Metal volumes lack paired clean reconstructions, so they support only qualitative and no-reference evaluation.Real-case metrics describe changes in artifact-like regions and smoothing across slices, not full-reference image quality.Source: C. Real CLINIC-Metal Validation · p. 4

D. Comparators and Ablations

  • Comparators include classical filters, pseudo-sinogram methods, and public pretrained learned models treated as deployment-transfer tests, with correction parameters fixed.Exact-mask oracle experiments isolate correction under accurate localization; automatic HU localization assesses the heuristic end-to-end pipeline.Source: D. Comparators and Ablations · p. 4

E. Metrics and Statistics

  • Synthetic image metrics use a shared metal-excluded tissue ROI, with 95% bootstrap confidence intervals, paired Wilcoxon tests, and Holm correction across nine contrasts.For unpaired real cases, input-relative metrics describe correction behavior only.Source: E. Metrics and Statistics · pp. 4–5

F. External Stress Test

On held-out synthetic cases, VoxelSynth3D reduced metal-excluded tissue RMSE relative to input and 3D Gaussian filtering, while clean-edge analysis showed a localized tradeoff. The external AAPM test was a reconstructed-image stress test, not native 3D or clinical validation.

F. External Stress Test

  • VoxelSynth3D was applied without parameter adjustment to 29 public AAPM CT-MAR images, replicated across five isotropic slices, as an external reconstructed-image stress test.Reference images and scoring masks were withheld during inference, mask construction, and adaptation; the test was neither native 3D nor clinical validation.Source: F. External Stress Test · p. 5

V. RESULTS AND DISCUSSION

  • The evaluation frames three questions: whether error reduction persists after metal exclusion, how structural cost varies with distance, and how mask extent affects target scores and unaffected tissue.These analyses precede external and real-data evaluations.Source: §V · p. 5, A. Full-Reference Synthetic Evaluation · p. 5

A. Full-Reference Synthetic Evaluation

B. Ablation, Severity, and Metric Tradeoff

Ablations attribute most observed full-ROI improvement to the volumetric prior, deviation gating, and adaptive support, while spatial analysis shows a near-metal edge cost alongside better edge agreement farther away.

B. Ablation, Severity, and Metric Tradeoff

  • Most observed full-ROI improvement came from the volumetric prior, deviation gating, and adaptive support; edge-guided refinement added a small endpoint improvement.VoxelSynth3D-MS improved full-ROI EPI to 0.559 from 0.542 but worsened RMSE to 785.72 from 785.43 HU and nearly doubled runtime; high artifact-size cases had 9.67% full-ROI MSE reduction versus 3.42% for pseudo LI-MAR and 3.49% for InDuDoNet.Source: B. Ablation, Severity, and Metric… · p. 6, Figure 5 · p. 6
  • At 2–5 mm from metal, mean RMSE fell from 858.90 to 244.84 HU while mean clean-edge F1 fell from 0.548 to 0.352; at 5–10 and 10–20 mm, F1 exceeded input by 0.089 and 0.132.Edge loss was concentrated where correction was strongest, and the lower RMSE does not establish edge recovery.Source: B. Ablation, Severity, and Metric… · p. 6

C. Mask Dependence and Collateral Modification

Mask quality limits the interpretation of correction gains: broader automatic localization can improve target-ROI RMSE while changing more unaffected tissue. Validation and qualitative comparisons remain bounded by the available evidence, especially for real cases and external stress tests.

C. Mask Dependence and Collateral Modification

  • Automatic HU localization improved target-ROI RMSE to 767.99 HU at precision 0.285, while outside-ROI RMSE rose by 39.22 HU.Only 0.082% of outside-ROI voxels changed by more than 50 HU with the exact-mask oracle, versus 1.04% with automatic localization.Source: C. Mask Dependence and Collateral… · p. 6

D. Validation Stability and External/Real-Data Scope

  • The reported setting (85, 1.0 mm, 0.45; 779.39 HU) was 0.10 HU above the validation minimum and differed by 0.09 HU on test.The same full-ROI trend persisted across metal-burden strata.Source: D. Validation Stability and… · p. 6
  • The external AAPM stress-test score changed from 2.2700 to 2.4091, but its case-level 95% CI crossed zero (p = 0.381), with 12/29 images improving.Repeating each 2D image into five identical slices makes this a reconstructed-image stress test, not native volumetric, leaderboard, or clinical evidence.Source: D. Validation Stability and… · p. 6
  • Real CLINIC-Metal cases have no clean targets, so their comparisons provide qualitative/no-reference evidence only, not evidence of anatomical correctness or diagnostic benefit.The paper calls for reader- and task-based validation for diagnostic claims.Source: D. Validation Stability and… · pp. 6–7

E. Qualitative Assessment

  • Figure 6 presents two objectively selected held-out cases with large tissue-RMSE gains over shown comparators, while Figure 7 provides qualitative/no-reference real-case comparisons.Figure 6 shows clean targets and metal-excluded absolute-error maps, where darker colors indicate lower error and cyan outlines the implant.Source: E. Qualitative Assessment · p. 7, Figure 6 · p. 7, Figure 7 · p. 7

F. Limitations

The paper’s evidence is limited by the synthetic pelvis evaluation scope, reduced near-metal robustness, and the lack of clinical validation. Its conclusion frames the framework as a within-distribution correction approach rather than a replacement for scanner-native MAR.

F. Limitations

  • Synthetic artifacts may not capture every scanner, protocol, implant, or reconstruction-kernel effect, and the pelvis-focused resource does not establish generalization to other anatomies or implant geometries.The spacing-corrected artifact model retained broad-region gains but showed reduced near-metal robustness.Source: F. Limitations · p. 7
  • Blinded radiologist scoring remains the key clinical validation step; AAPM and real input-relative metrics cannot establish diagnostic benefit.Source: F. Limitations · p. 7

VI. CONCLUSION

  • VoxelSynth3D’s evidence supports measurable within-distribution benefits with limitations, not replacement of scanner-native MAR; calibrated simulation and blinded task assessment are needed for external and clinical effectiveness.Source: §VI · p. 7
Loading 2610.01512v1…