Mitigating Metal Artifacts in Medical Imaging

427 concepts303 papers12 proposals

Medical imaging modalities rely on the predictable interaction of energy—whether X-ray photons or radiofrequency pulses—with human tissue. However, when a high-density metallic object such as a hip prosthetic, dental filling, or deep brain stimulation lead is introduced, these physical assumptions break down. The result is a suite of severe distortions known collectively as metal artifacts, which obscure critical anatomy and degrade the performance of downstream diagnostic algorithms.

To solve this problem, researchers attack the issue from multiple angles: rewriting the underlying physics of image reconstruction, using generative deep learning to predict missing data, altering hardware environments, and designing computer vision models that are inherently invariant to noise.

The Physics of Metal Interference

To understand why metal degrades an image, we must look at how different modalities capture data.

In Computed Tomography (CT), X-rays pass through the body to detectors on the opposite side. When these rays encounter high-density metal, two primary phenomena occur: beam hardening and photon starvation. X-ray beams are polychromatic, meaning they contain a spectrum of energies. As they pass through metal, lower-energy photons are preferentially absorbed, shifting the mean energy of the remaining beam ("hardening" it). Simultaneously, so few photons manage to penetrate dense metal that the detectors experience "photon starvation." When a reconstruction algorithm tries to back-project these incomplete signals, it creates characteristic dark and bright streaks across the image, obscuring bone and tissue boundaries as described by Neslisah Torosdagli et al.

In Magnetic Resonance Imaging (MRI), the problem stems from magnetic susceptibility. Human tissue is weakly diamagnetic, but metallic implants are strongly paramagnetic or ferromagnetic. This creates a massive mismatch in the local magnetic field ($B_0$). Hydrogen protons near the metal experience a highly distorted field, causing them to precess at unexpected frequencies. This leads to rapid phase dephasing—manifesting as large black "signal voids"—and spatial mismapping, where signals are reconstructed in the wrong place, causing bright "blooming" artifacts (Yuli Wang et al.).

Advanced Reconstruction and Signal Processing

One way to mitigate these artifacts is to modify the mathematical framework used during image reconstruction. Traditional methods often fail because they treat all projection data equally, regardless of whether it passed through metal.

To address this, Peng Bao et al. introduce a sparse-view CT reconstruction method using Convolutional Sparse Coding (CSC). By regularizing feature maps and enforcing data-fidelity constraints, this approach suppresses high-frequency streak artifacts while preserving fine anatomical details that traditional smoothing filters would blur away.

In dual-energy CT (DECT), the authors of a study on shoulder arthrography (Christoph Stern et al.) leverage virtual non-contrast (VNC) imaging. By mathematically subtracting high-density contrast agents based on their energy-attenuation profiles, they prevent the contrast material from compounding pre-existing metal-induced distortions.

However, algorithmic fixes require careful calibration. Robert Hatherly et al. caution that standard software-based metal artifact reduction (MAR) algorithms can introduce secondary errors. In functional PET/CT imaging, these algorithms can misinterpret attenuation data near high-density hardware, creating artificial "hot spots" or false-positive signals. In certain high-interference environments, omitting these specific software corrections entirely can actually yield more reliable metabolic maps.

Deep Learning and Generative Inpainting

When physical data is completely lost to photon starvation or signal voids, traditional signal processing cannot recover it. Deep learning models can bridge this gap by learning the underlying structural priors of human anatomy and predicting what the obscured regions should look like.

Dongming Wei et al. tackle the severe streak artifacts caused by metallic ablation probes during liver tumor procedures. They combine cross-modality image synthesis (using a Cycle-GAN to map between MRI and CT spaces) with generative image inpainting. The network identifies the corrupted streak regions, digitally erases them, and "inpaints" the missing anatomy, enabling highly accurate automated image registration.

[Artifact-Heavy Scan] ──> [Identify Corrupted Streaks] ──> [Generative Inpainting] ──> [Artifact-Free Anatomy]

Similarly, Hana Baroudi et al. use Generative Adversarial Networks (GANs) to synthesize megavoltage (MV) cone-beam CT images from lower-quality kilovoltage (kV) scans. Because MV photons interact differently with high-density matter, this synthesis bypasses the severe beam-hardening and scattering artifacts that typically obscure the boundaries of cardiac pacemakers during radiation therapy planning.

Hardware and Material Innovations

Rather than correcting artifacts post-acquisition, a parallel line of research aims to prevent them entirely by altering the physical imaging environment or the composition of the implants themselves.

In MRI, magnetic field distortion scales with the strength of the main magnetic field. Juliet Varghese et al. demonstrate that modern low-field (0.55T) MRI systems are inherently more resilient to field inhomogeneities than standard 1.5T or 3T scanners. This lower field strength drastically reduces susceptibility artifacts, making it possible to acquire diagnostic-quality cardiac images in patients with pacemakers or stents without severe signal dropouts.

Alternatively, the materials used inside the body can be redesigned to be "imaging-visible" without being disruptive. In a single-author study, Mingming Yin developed a carbon-iodine conjugated polymer to replace traditional metallic tumor-tracking fiducial markers, eliminating CT streak artifacts entirely. For interventional procedures, Liam Cunningham et al. designed 3D-printed catheters impregnated with liquid vitamin D. This provides passive, artifact-free visibility under MRI without relying on metallic markers or electronic coils that warp the local magnetic field.

Designing for Downstream Robustness

A final strategy accepts that artifacts will occur and instead focuses on building computer vision models that are invariant to high-contrast noise. Traditional segmentation algorithms often fail near metal because they rely on local voxel intensities; when streaks or blooming alter these values, the algorithm loses the boundary.

To circumvent this, Hyojin Kim et al. demonstrate that a randomized ensemble of simple segmentation models can successfully delineate objects in industrial CT scans corrupted by cupping and blooming artifacts. Instead of relying on a single "perfect" parameter set, the ensemble averages out artifact-induced edge errors, providing a more robust boundary estimation.

In clinical anatomy, Neslisah Torosdagli et al. developed a multi-stage pipeline that combines Random Forest regression with fuzzy logic to segment the mandible from scans cluttered with heavy dental hardware artifacts. By incorporating an anatomical "state-machine" that understands the global geometric relationships of the jaw, the system prevents local streaking noise from pulling the segmentation boundaries off course.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 299 papers
  2. Version 1 · 2026-10-04 · 217 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_metal_artifact_interference_e58b0966,
  title = {metal artifact interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/metal-artifact-interference/rdr_e58b0966a01e4d49beb3bc7600839933},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_metal_artifact_interference_e58b0966_v2,
  title = {metal artifact interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/metal-artifact-interference/rdr_e58b0966a01e4d49beb3bc7600839933/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake