Metal Artifact Reduction (MAR)

The foundational approach where corrupted projection data in the sinogram domain is treated as missing and replaced via interpolation.

• Mechanism: Metallic objects cause photon starvation, creating null or near-null values in the raw detector readings. MAR algorithms segment the metal trace in the sinogram and replace it with estimated values from surrounding uncorrupted projections.
• Common Techniques: Linear interpolation, directional interpolation, and normalized metal artifact reduction (NMAR) which uses a prior image to guide the inpainting.
• Limitation: Simple interpolation often introduces new secondary artifacts due to inconsistencies at the boundary between original and interpolated data.

Modern model-based iterative reconstruction (MBIR) algorithms integrate MAR directly into the reconstruction loop by down-weighting or discarding unreliable projection data.

• Statistical Weighting: Each X-ray projection is assigned a reliability weight based on its noise variance. Rays passing through metal receive a weight near zero, effectively excluding them from the objective function.
• Regularization: Edge-preserving regularization terms, such as total variation (TV) minimization, penalize the streaking patterns characteristic of metal artifacts while preserving anatomical boundaries.
• Advantage: This approach avoids the secondary artifacts common in sinogram inpainting by solving a single, consistent optimization problem.

Convolutional neural networks are trained end-to-end to map artifact-corrupted images directly to artifact-free images, bypassing explicit sinogram processing.

• Supervised Training: Networks like ResNet or U-Net variants are trained on paired datasets where the same anatomy is imaged with and without metal, or where artifacts are simulated from clean images.
• Dual-Domain Networks: Advanced architectures operate simultaneously on both the sinogram and image domains, learning to correct inconsistencies in projection space and refine structural details in image space.
• Generative Approaches: GANs and diffusion models are employed to hallucinate plausible tissue textures in regions where the original signal is completely lost behind dense metal.

A seminal algorithmic framework that combines tissue segmentation, forward projection, and normalization to create a high-fidelity prior image for sinogram inpainting.

• Process: The original corrupted image is segmented into tissue classes (air, soft tissue, bone). A prior image is created by re-projecting this segmentation. The original sinogram is normalized by the prior sinogram, inpainted, and then de-normalized before final reconstruction.
• Key Insight: Normalization transforms the high-dynamic-range projection data into a flat, low-frequency space where interpolation errors are dramatically reduced.
• Clinical Impact: NMAR preserves fine bone-tissue interfaces near prostheses that are often blurred by simpler interpolation methods.

A hybrid strategy that separates the image into low-frequency and high-frequency components, applying different correction strategies to each.

• Rationale: Metal artifacts manifest as high-contrast streaks (high-frequency) and broad dark bands (low-frequency). Treating them separately prevents over-smoothing of anatomical edges.
• Implementation: The low-frequency component, containing the dark shading artifacts, is corrected using a tissue-class model or prior image. The high-frequency component, containing edges and noise, is preserved or adaptively filtered.
• FS-MAR: The Frequency Split Metal Artifact Reduction algorithm applies this principle to avoid the loss of spatial resolution common in aggressive sinogram interpolation.

A technique specific to dual-energy CT (DECT) that synthesizes images at high keV levels to physically suppress beam hardening, the root cause of many metal artifacts.

• Principle: Beam hardening occurs because low-energy photons are preferentially absorbed by dense metal. By computationally synthesizing a monochromatic beam at high energy (e.g., 120-200 keV), the polychromatic artifact source is eliminated.
• Clinical Use: Radiologists routinely review high-keV virtual monoenergetic reconstructions to assess the bone-prosthesis interface and periprosthetic soft tissues.
• Synergy: VMI is often combined with iterative reconstruction or deep learning MAR for additive artifact suppression.

The classic analytic CT reconstruction algorithm that applies a high-pass ramp filter to projection data before back-projecting across the image grid. Metallic implants violate FBP's fundamental assumption of monochromatic, noise-free projections, causing the severe streaking artifacts that MAR algorithms are designed to correct. Modern MAR techniques often operate in the projection domain to repair the sinogram data before FBP is applied.

A computationally intensive reconstruction framework that repeatedly compares forward-projected model estimates with raw acquisition data to converge on a noise-optimized image. Unlike FBP, IR can incorporate physical models of photon starvation and beam hardening directly into the reconstruction loop, allowing it to down-weight corrupted projection data caused by metal. Model-based IR variants are inherently more robust to metal artifacts than analytic methods.

A physical phenomenon where a polychromatic X-ray beam becomes progressively more energetic as lower-energy photons are preferentially absorbed by dense material. This shifts the effective attenuation coefficient and produces characteristic cupping artifacts and dark bands between metallic objects. MAR algorithms must explicitly model or compensate for this non-linear spectral effect, often through polynomial correction or dual-energy decomposition.

Occurs when metallic implants absorb nearly all incident X-ray photons along certain projection angles, resulting in zero or near-zero detector measurements dominated by electronic noise. The log-transform in standard CT reconstruction amplifies this noise exponentially, creating the characteristic bright and dark streaking radiating from metal. MAR strategies address this by interpolating across corrupted detector channels or using normalized sinogram inpainting.

A modern class of reconstruction algorithms using convolutional neural networks trained on paired low-quality and high-quality CT data to suppress noise and resolve fine structures. For MAR, DLR models can operate in the projection domain (sinogram inpainting), the image domain (artifact-to-clean mapping), or a dual-domain approach that corrects both. These networks learn to distinguish metal-induced streaks from genuine anatomical edges.

A seminal algorithmic framework that reduces the impact of interpolation errors by first creating a prior image (often a thresholded, metal-free initial reconstruction), forward-projecting it, and performing sinogram inpainting in the normalized projection domain. By dividing the original sinogram by the prior sinogram before interpolation, NMAR preserves fine anatomical structures near metal that would otherwise be smoothed away by direct interpolation.