Home›Articles›CT
← Back to Articles

Metal Artifact Reduction in CT: A Guide for Radiographers

AI-assisted educational writing · Radiography 101
Editorial policy · No qualified clinical review claimed

Metal in the scan field is one of the few CT problems a radiographer can see coming before the patient is on the table — a hip replacement, spinal fixation, dental work, a pacemaker, a coil mass — and one that cannot be fully solved afterwards. What you can change is how much of the resulting artifact reaches the images the radiologist reads.

This guide separates the physics from the marketing. It explains what metal actually does to the data, what acquisition choices and reconstruction algorithms each do about it, what the current human studies measured, and where the tradeoffs land. It deliberately does not give you a tube voltage or a keV value to apply to every implant, because the published values come from specific protocols on specific hardware.

Short answer

Metal produces streaks through three separate mechanisms — beam hardening, photon starvation, and scatter — and no single setting addresses all three. Reconstruction-based metal artifact reduction (MAR) such as iterative MAR attacks the raw data and the image together; high-keV virtual monoenergetic images (VMI) reduce beam hardening only, and mainly when artifacts are mild. In the comparative studies cited here, iMAR-based reconstructions scored highest, and combining a high-keV VMI with iMAR produced the highest median scores of all — although that gain over iMAR alone was not statistically significant. Every correction also removes something real, so the uncorrected series should stay available for review.

Why metal breaks the image

Three mechanisms act at once, and they are worth keeping separate because they respond to different interventions.

Beam hardening. A CT tube emits a spectrum, not a single energy. Low-energy photons are attenuated more strongly than high-energy ones, so the beam becomes progressively "harder" as it passes through material. Metal hardens the beam far more than soft tissue does in the same path length, and the reconstruction assumes a monochromatic beam. The mismatch appears as dark bands and cupping between dense structures.

Photon starvation. Very few photons emerge from the far side of a metal implant. Where the detector receives a small number of photons, the statistical uncertainty on that measurement is large, and the reconstruction amplifies it. The visible result is high noise and streaks aligned with the paths that were starved.

Scatter. Photons that are scattered rather than absorbed can reach detector elements belonging to a different ray. The reconstruction then assigns that extra signal to the wrong line integral, which is one reason artifacts spill well beyond the implant itself.

Schematic showing three X-ray paths through a patient in CT. Path A crosses soft tissue only and delivers the expected photon count. Path B crosses a titanium implant and delivers very few photons, illustrating photon starvation. Path C runs beside the metal and receives scattered photons from a path it did not travel. The right panel shows the resulting reconstruction with bright streaks, dark beam-hardening bands and obscured soft tissue around the implant, and a legend below explains the three paths.
Figure 1 — Why metal breaks a CT reconstruction. The three mechanisms act on the same ray paths at the same time, which is why a single correction rarely clears the whole artifact. Original diagram for this article; not a clinical image.

For a broader map of artifact patterns and how to tell them apart at the console, the companion guide to CT artifacts and their causes covers motion, rings, partial volume and noise alongside metal. This article goes deeper on metal specifically.

Acquisition-side options and their limits

Some control exists before reconstruction. A published comparison of MAR techniques notes that metal artifacts can be reduced by increasing tube voltage or by repositioning the patient.2 Both work by changing the physics rather than by post-processing: a harder beam is less distorted by the implant, and moving the implant out of the most critical part of the field changes which anatomy the artifacts fall across.

Neither is free. Raising tube voltage moves the spectrum away from the photoelectric region where soft-tissue contrast is generated, and it changes dose. Repositioning is not always possible in a trauma patient or in a fixed protocol. These are protocol decisions for the radiologist and medical physicist, not console improvisations.

Practical acquisition habits still matter and cost nothing in dose: keep the implant out of the reconstruction centre where you can, collimate to the anatomy actually needed, avoid repeating a scan to "see if the artifact is better" without a change that would plausibly achieve it, and record what reconstruction options were used so the next scan of the same patient is comparable.

Acquisition versus iterative MAR

The phrase "MAR" covers techniques that act at completely different points in the data chain, and lumping them together causes most of the confusion in this topic.

Four-stage pipeline from X-ray tube and patient to raw projection data, reconstruction, and the image sent to PACS. Bands underneath show acquisition choices acting at the tube and patient stage, iterative metal artifact reduction acting on raw data and in the image domain, high-keV virtual monoenergetic imaging requiring spectral data at reconstruction, and deep-learning reduction operating on reconstructed images. A highlighted note at the bottom says to keep the uncorrected reconstruction alongside the corrected one because correction can create pseudo-osteolysis.
Figure 2 — Where each approach acts. Different stages, different failure modes. A tool that fixes the streaks is not automatically a tool that preserves the bone detail next to them. Original diagram for this article.

Iterative MAR works in two places at once

A commercial iterative MAR implementation, as described in the photon-counting study cited below, addresses beam hardening by correction and inpainting in raw data space using a normalised sinogram. It then applies a frequency-split technique in the image domain to reduce what remains.1 That two-stage structure is why iterative MAR can outperform a purely spectral correction: it is working on the corrupted measurements, not only on the reconstructed result.

The same source notes that dedicated iterative MAR algorithms are typically provided per implant type — hip, intracranial coils, thoracic coils, shoulder, pacemakers, dental fillings, extremity implants — and that a hip algorithm may be applied to a region with no dedicated option, such as spinal hardware, because nothing better exists. That is a real limitation, not a footnote: the algorithm's assumptions were not designed for that geometry, and the images should be read with that in mind.

High-keV virtual monoenergetic images are a different tool

VMI requires spectral data, from dual-energy CT or from a photon-counting system. Reconstructing at a high virtual energy reduces beam hardening because the reconstruction behaves as though the beam were monochromatic at that energy, so the low-energy photons that harden the beam are no longer represented.

The evidence puts a ceiling on this. A total hip arthroplasty study comparing techniques describes high-energy VMI as useful for relatively mild artifacts but of limited use when artifacts are severe, and notes that the optimal energy depends on implant material, size and shape — while 130 keV was generally effective for the hip prostheses in the cited work.2 A separate photon-counting study reconstructed at 140 keV and found the benefit was significant mainly for distant anatomy, with only minor artifact reduction adjacent to a hip replacement; the same 140 keV reconstruction performed substantially better around spinal osteosynthesis.1

That difference between implant types in the same study is the single most useful thing to take from this section. The keV value is not a setting you carry between patients.

Where photon-counting CT fits

Photon-counting detectors register individual photons and resolve their energy, so spectral data is available from every acquisition rather than requiring an up-front decision to scan in a dual-energy mode.1 In the context of metal, high-energy bin images show fewer beam artifacts than low-energy bin images or conventional energy-integrating CT, and the data set can be reconstructed at high keV retrospectively even if no one planned for it.

The wider musculoskeletal review cited here sets out the same advantages — better spatial resolution, reduced noise and intrinsic spectral capability — and adds the mechanism that matters for metal: spectral information allows beam-hardening and metal artifacts to be reduced without increasing radiation exposure.3 The same review also shows the practical catch, and it is the most important image in this article's source set: iterative MAR improves soft-tissue assessment around a total hip replacement, but the correction can create appearances mimicking areas of osteolysis, so bone structures are best assessed on the reconstruction without the algorithm.3

What that means at the console: a photon-counting scanner does not remove the need for a reconstruction strategy. It makes more reconstructions available, which increases the value of labelling them accurately and of keeping at least one algorithmic correction out of the bone window.

What the human evidence actually shows

Three peer-reviewed papers are relevant enough to describe individually: two small single-centre reader studies and one review. Each supports a statement about the conditions studied rather than a general rule.

The cited studies compared techniques within a defined implant population. Sample sizes and reader numbers are reported as published.
StudyPopulationWhat was comparedMain finding
Pallasch et al., Radiol Med 20241 39 patients with orthopaedic implants: 13 spinal osteosynthesis, 13 hip replacement, 9 other osteosynthesis, 2 knee, 2 shoulder Photon-counting CT reconstructed four ways: clinical standard, 140 keV VMI, iMAR, and 140 keV + iMAR; three blinded radiologists Highest scores for 140 keV + iMAR across every criterion, then iMAR alone, then the non-iMAR reconstructions
Selles et al., Eur Radiol Exp 20242 28 patients with unilateral total hip arthroplasty Conventional CT, 130 keV VMI, each with and without O-MAR, plus a deep-learning MAR; two musculoskeletal radiologists Deep-learning MAR scored higher than conventional and 130 keV images with O-MAR for image quality, artifact severity and confidence in bone and adjacent soft tissue
Mourad et al., Skeletal Radiol 20243 Review of photon-counting CT in musculoskeletal imaging Technical features against clinical applications, including metal hardware Spectral data reduces beam-hardening and metal artifacts without raising exposure, but iterative correction can mimic osteolysis

The numbers behind the first row

In the photon-counting study, median diagnostic confidence rose from 2 out of 5 on the clinical standard reconstruction to 3 at 140 keV, 3.7 with iMAR and 4 with the combination. The combination was significantly better than the standard for every criterion, iMAR alone was significant for all criteria except diagnostic confidence, and 140 keV alone reached significance only for delineation of distant anatomy.1

One negative result in that study is worth as much as the positive ones: the coefficient of variation measured in unaffected muscle did not differ significantly between reconstruction approaches. In other words, the algorithms changed the artifact significantly without changing measured signal homogeneity in tissue away from the implant. If you expect a corrected series to look "cleaner" everywhere, that is not what the measurement showed.

And the deep-learning comparison

In the hip arthroplasty study, deep-learning MAR scored higher than both conventional and 130 keV images with O-MAR for image quality, artifact severity, and confidence in bone and in soft tissue adjacent to the prosthesis, with higher contrast-to-noise ratios.2 The exception is instructive: confidence in pelvic organs was not significantly different from conventional CT with O-MAR. The algorithm's advantage was local to the prosthesis, not global.

That study also reported that the deep-learning algorithm was trained on simulated data — over 105,000 paired images — which is the usual development route and also the usual caveat. Good performance on clinical images in one centre is the beginning of external validation, not the end of it.

Tradeoffs you have to accept

Reduced contrast. High-keV reconstructions lose overall image contrast as a direct consequence of moving away from the energies where soft-tissue differences are largest.2 A smoother-looking image is not automatically a more diagnostic one, particularly in contrast-enhanced studies.

False defects. This is the tradeoff that changes reporting. Iterative correction can create appearances that mimic osteolysis around hardware, which is why bone is best assessed on the reconstruction without the algorithm.3 A radiographer who sends only the corrected series has made a diagnostic decision by accident.

Limited benefit where artifacts are worst. High-keV VMI is described as most useful for relatively mild artifacts and of limited use when artifacts are severe, where reconstruction-based MAR is the tool of choice.2

Algorithm mismatch. Dedicated algorithms are built per implant type, and regions without one are handled by borrowing a neighbouring algorithm.1 Expect that borrowed behaviour to be less predictable.

Small evidence base. Thirty-nine patients is enough to show a consistent reader preference. It is not enough to tell you how the technique performs across implant materials, body sizes, or the artifacts that matter most in your department. The musculoskeletal review closes by calling for further preclinical and clinical validation of these applications.3

Why the uncorrected reconstruction travels with the study

If MAR can imitate pathology, the uncorrected series is not a redundant image — it is the reference. The practical rule that follows from the sources above is simple: reconstruct both, label both unambiguously, and let the reporting radiologist compare. Whether the uncorrected series is sent to PACS routinely, or retained on the scanner for review, is a departmental decision. That it exists is not optional.

Labelling deserves the same care as acquisition. An unnamed series that "looks different" is a hazard in a comparison study, and reconstruct naming that does not state the keV, the algorithm, and whether the algorithm was applied becomes unusable once the patient returns six months later.

What this means for orthopaedic hardware in practice

Error-to-correction exercise

Claim to examine: "Our scanner has metal artifact reduction, so metal cases are solved."

What is wrong with it: it treats MAR as one thing. Iterative MAR, high-keV VMI and deep-learning MAR act at different stages, address different mechanisms, and fail differently. In the photon-counting study, 140 keV alone reached significance only for delineation of distant anatomy, while iMAR-based reconstructions were significantly better across nearly every criterion.1 In the hip study, the deep-learning algorithm's advantage did not extend to confidence in pelvic organs.2

Correction: "Our scanner offers several MAR options that act at different points in the data chain. Which one helps depends on the implant and the clinical question, the correction can mimic pathology near bone, and the uncorrected reconstruction stays available for comparison."

Key takeaways

  1. Beam hardening, photon starvation and scatter are three distinct mechanisms; correcting one does not correct the others.
  2. Iterative MAR works on raw data and in the image domain; high-keV VMI is a spectral reconstruction that mainly helps mild artifacts.
  3. In a 39-patient photon-counting study, 140 keV plus iMAR scored highest, iMAR alone second, and 140 keV alone helped mainly distant anatomy.
  4. Correction can imitate osteolysis, so bone is assessed on the uncorrected reconstruction and both series should travel together with clear labels.
  5. Published keV values reflect specific protocols and implants; there is no universal setting, and no dose increase is justified by artifact alone.

Where to go next

If you want the detector physics behind the spectral options described here, photon-counting CT explained for radiographers walks through how energy-resolved detection differs from a conventional scintillator detector. For the reconstruction fundamentals that sit underneath every artifact discussion, see CT scan physics. The CT modality page collects the related guides, and the free practice test centre includes physics and image-production questions if you are revising for a registry exam.

References

  1. Pallasch FB, Rau A, Reisert M, et al. Photon-counting detector computed tomography for metal artifact reduction: a comparative study of different artifact reduction techniques in patients with orthopedic implants. La Radiologia Medica. 2024;129(6):890–900. doi:10.1007/s11547-024-01822-x
  2. Selles M, Wellenberg RHH, Slotman DJ, et al. Image quality and metal artifact reduction in total hip arthroplasty CT: deep learning-based algorithm versus virtual monoenergetic imaging and orthopedic metal artifact reduction. European Radiology Experimental. 2024;8:31. doi:10.1186/s41747-024-00427-3
  3. Mourad C, Gallego Manzano L, Viry A, et al. Chances and challenges of photon-counting CT in musculoskeletal imaging. Skeletal Radiology. 2024;53(9):1889–1902. doi:10.1007/s00256-024-04622-6
Educational note: This article explains reconstruction technology for learning purposes and reports results under the conditions each study described. It is not a protocol. Local protocols, vendor training, radiologist direction and medical-physics guidance govern clinical scanning decisions, and published keV values should not be applied to a patient without local validation.
Review status: This is an AI-assisted educational draft. It has not been reviewed by a qualified radiologist or medical physicist, and no such review is claimed. A qualified reviewer should check the physics descriptions, the study summaries and the practical statements before publication.