⚡ electrical engineering

MGMAR: Metal-Guided Metal Artifact Reduction for X-ray Computed Tomography

The paper proposes MGMAR, a state-of-the-art metal artifact reduction method for CT imaging that integrates a pretrained, metal-guided implicit neural representation for robust projection completion with a metal-conditioned correction network to effectively suppress severe streaking artifacts and preserve anatomical structures.

Original authors: Hyoung Suk Park, Kiwan Jeon

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hyoung Suk Park, Kiwan Jeon

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Metal Glitch" in Medical Scans

Imagine you are trying to take a clear photograph of a room using a flashlight. Most of the room is visible. But, suddenly, you shine the light directly at a giant, shiny mirror or a thick block of lead.

In the world of medical CT scans (which are like taking thousands of X-ray photos from different angles to build a 3D model), metal implants (like hip replacements, dental fillings, or spinal screws) act like that giant mirror or block of lead.

When the X-rays hit the metal, they get confused. They scatter, they get "hardened" (losing energy), and they starve the detectors of light. When the computer tries to rebuild the image, it gets the math wrong. The result? The final image looks like a TV screen with bad reception: streaks of white and black lines shooting out from the metal, and dark shadows hiding the important organs nearby. Doctors can't see tumors or fractures through this "static."

The Solution: MGMAR (The "Metal Detective")

The authors propose a new method called MGMAR (Metal-Guided Metal Artifact Reduction). Instead of just trying to fix the picture after it's blurry, MGMAR acts like a detective that knows exactly where the metal is and uses that knowledge to guide the reconstruction.

Think of MGMAR as a three-step process to clean up a muddy, distorted photo:

Step 1: The "Smart Sketch" (The Prior Image)

Before trying to fix the messy scan, the computer needs a "best guess" of what the clean image should look like.

Old Way: The computer would start with a blank canvas and try to guess the image from scratch. If the metal was huge, it would get confused and produce a different, messy sketch every time you ran it.
MGMAR Way: The computer has a "training camp." It has seen thousands of examples of "dirty" scans and their "clean" versions. It learns a data-driven shortcut.
- The Analogy: Imagine an artist who has practiced drawing faces so much that they can instantly sketch a rough outline of a face without looking at a reference. MGMAR uses this "muscle memory" to create a high-quality sketch (called a Prior Image) of the patient's anatomy, ignoring the metal for a moment.
- The Secret Sauce: To make this sketch even better, the computer looks at the "ghost" of the metal artifacts (the streaks) and uses that as a clue to understand the global pattern of the distortion. It's like looking at the smoke to figure out where the fire is, so you can draw the fire correctly later.

Step 2: The "Targeted Repair" (NMAR)

Now that we have a good sketch, we need to merge it with the actual X-ray data.

The Process: The computer takes the messy X-ray data and the "Smart Sketch." It normalizes them (makes them comparable) and then performs a "projection completion."
The Analogy: Imagine the X-ray data is a torn map. The metal has ripped out the most important parts. The "Smart Sketch" is a perfect copy of the map you have in your pocket. MGMAR takes the torn map and carefully fills in the missing holes using the perfect copy, but only in the ripped areas. It leaves the rest of the map (the healthy tissue) exactly as the X-ray saw it, preserving the sharp details.

Step 3: The "Fine-Tuning" (Residual Correction)

Even after Step 2, there might be some leftover smudges or weird streaks near the metal because the "filling" wasn't 100% perfect.

The Process: MGMAR uses a special neural network to look for these remaining errors.
The Analogy: Think of this as a photo editor with a magic brush. But this brush is "metal-aware."
- If you use a normal editor, they might accidentally blur the patient's nose while trying to fix a smudge near their ear.
- MGMAR's editor has a mask (a stencil) that covers the healthy parts of the face. The brush only touches the metal area and the immediate neighborhood. It specifically targets the "secondary artifacts" (the leftover streaks) without blurring the healthy bones or organs.

Why is this better than what we had before?

It's Stable: Old methods were like rolling dice; sometimes they worked, sometimes they didn't. MGMAR uses its "training camp" (pre-training) to start with a solid foundation, so it always produces a reliable result.
It's Smart: It doesn't just guess; it uses the specific shape of the metal artifacts to guide the correction.
It Wins: When tested on real clinical cases (the AAPM-MAR challenge), MGMAR scored the highest. It produced the clearest images, the least noise, and the most accurate bone structures compared to all other top methods.

The Bottom Line

MGMAR is like having a super-smart, metal-savvy artist who can look at a ruined, streaky X-ray and reconstruct a crystal-clear picture of the inside of the body. It does this by:

Learning from past examples to make a rough sketch.
Filling in the missing X-ray data using that sketch.
Polishing the final image to remove any last smudges, all while being careful not to touch the healthy parts.

This helps doctors see clearly through the "static" caused by metal implants, leading to better diagnoses and safer treatments.

1. Problem Statement

In X-ray Computed Tomography (CT), metallic implants (e.g., hip prostheses, dental fillings, spinal screws) violate the linear attenuation assumptions of standard reconstruction algorithms like Filtered Back-Projection (FBP). This violation causes severe metal artifacts, primarily manifested as streaking and shadowing, which degrade diagnostic quality.

Existing Metal Artifact Reduction (MAR) techniques face specific challenges:

Projection Completion Methods (e.g., NMAR): Rely on a "prior image" to normalize sinograms. If the prior image is inaccurate, the interpolation of corrupted data introduces secondary artifacts.
Model-Based Methods: Require detailed knowledge of the X-ray spectrum and material properties, leading to high computational costs.
Deep Learning Approaches: While effective, many struggle with stability when metal-unaffected projections are severely limited by large implants. They often suffer from sensitivity to random initialization and fail to capture global artifact patterns.

2. Methodology: MGMAR Framework

The authors propose MGMAR (Metal-Guided Metal Artifact Reduction), a framework that explicitly leverages metal-related information throughout the entire reconstruction pipeline. The process consists of three main stages:

A. Data-Driven Prior Image Generation via Conditioned INR

Instead of using a standard FBP image as a prior, MGMAR generates a high-quality prior using an Implicit Neural Representation (INR).

Conditioned INR: The image is modeled as a continuous function $f_\theta(x, z)$ , where $x$ is the spatial coordinate and $z$ is a latent code.
Artifact-Aware Conditioning: A convolutional encoder ( $E_\psi$ ) takes a two-channel input: the metal-corrupted CT image ( $\mu$ ) and a recursively constructed metal-artifact image ( $\mu_{MA}$ ). This allows the latent field to capture global streaking and shadowing patterns.
Recursive Construction of $\mu_{MA}$ : Since the ground-truth artifact image is unknown during inference, it is estimated recursively. The algorithm iteratively refines the artifact estimate by comparing the current NMAR-corrected projection against the measured projection.
Data-Driven Weight Initialization: To solve the instability of optimizing INRs from random initialization (which leads to varying results), the authors pretrain the encoder and INR on paired datasets (corrupted vs. artifact-free). This embeds data-driven prior knowledge into the network weights, providing a robust starting point ( $\theta_0$ ) for measurement-specific refinement.
Refinement: The pre-trained weights are fine-tuned on the specific patient's metal-unaffected projections to ensure data fidelity.

B. Normalized MAR (NMAR) Projection Completion

The high-quality prior image generated above is used within the standard NMAR framework:

The prior image is forward-projected to create a prior sinogram.
The measured sinogram is normalized by the prior sinogram.
Metal-corrupted regions in the normalized sinogram are interpolated using neighboring metal-free data.
The corrected sinogram is re-normalized and reconstructed via FBP.
Note: MGMAR simplifies the pipeline by removing redundant NMAR steps present in previous versions, relying on the superior quality of the INR-generated prior.

C. Metal-Embedded Residual Correction

Even after NMAR, secondary artifacts (residual streaks) often remain. MGMAR employs a residual learning network ( $G_\phi$ ) to suppress these:

Metal Conditioning via AdaIN: The network uses Adaptive Instance Normalization (AdaIN) to modulate intermediate feature maps based on the metal mask ( $\chi_D$ ).
Mechanism: A sub-network predicts affine parameters ( $\alpha, \beta$ ) from the metal mask, which are applied to the feature channels. This forces the network to focus its correction capacity specifically on metal-dependent artifacts while preserving non-metal anatomical structures.

3. Key Contributions

Metal-Guided Pipeline: A unified framework that explicitly integrates metal information (via masks and artifact images) at every stage: prior generation, projection completion, and residual correction.
Data-Driven INR Initialization: A novel strategy to pretrain INR weights on paired data, embedding global artifact patterns into the latent space. This significantly improves stability and convergence speed compared to random or meta-learning initializations.
Artifact-Aware Latent Conditioning: An encoder architecture that inputs both the corrupted image and a recursively estimated artifact map, enabling the capture of global artifact patterns essential for severe corruption cases.
AdaIN-Based Residual Correction: A mechanism that uses the metal mask to guide the residual network, ensuring corrections are targeted at artifacts without distorting healthy anatomy.

4. Experimental Results

The method was evaluated on the AAPM-MAR Grand Challenge dataset, comprising 14,000 simulated training cases and 29 clinical test cases.

Performance: MGMAR achieved a state-of-the-art final benchmark score of 0.89 (average of 8 metrics), outperforming the top 3 methods from the original challenge and the standard NMAR baseline.
Key Metrics: It achieved the best scores in CT Number (CTN) accuracy, Noise, SSIM, and Proton Beam Range (PBR) influence.
Ablation Studies:
- Initialization: Data-driven initialization outperformed random and meta-initializations, achieving near-optimal RMSE/PSNR even without further training iterations ( $N_{iter}=0$ ).
- Artifact Input: Including the recursively constructed artifact image ( $\mu_{MA}$ ) in the encoder input significantly improved prior image quality.
- Residual Conditioning: Metal conditioning (AdaIN) prevented the residual network from distorting anatomical structures, though the quantitative gain was modest due to the small magnitude of remaining artifacts.
Runtime: The total runtime is approximately 50 seconds per case (41s for INR refinement, 6s for prior generation, 2s for NMAR, 1s for residual correction).

5. Significance

MGMAR represents a significant advancement in CT metal artifact reduction by bridging the gap between physics-based reconstruction (NMAR) and deep learning.

Clinical Impact: By significantly reducing streaking and shadowing while preserving anatomical integrity, it improves diagnostic reliability for patients with metallic implants.
Methodological Innovation: The introduction of data-driven weight initialization for INRs addresses a critical bottleneck in continuous representation learning for ill-posed inverse problems, offering a more stable and efficient alternative to random initialization.
Future Directions: The authors note that while the method is effective, the reliance on paired training data (simulated) is a limitation for real-world deployment, and future work will focus on accelerating the refinement stage and extending the framework to 3D cone-beam CT.