MGMAR: Metal-Guided Metal Artifact Reduction for X-ray Computed Tomography
New AI method tackles streaking artifacts from implants, achieving state-of-the-art 0.89 score on clinical benchmark.
Researchers Hyoung Suk Park and Kiwan Jeon have introduced MGMAR (Metal-Guided Metal Artifact Reduction), a novel AI-powered method designed to solve a persistent problem in medical imaging: severe streaking and shadowing artifacts in X-ray computed tomography (CT) scans caused by metallic implants like screws or joint replacements. These artifacts violate standard CT model assumptions and degrade diagnostic quality. The proposed system explicitly leverages metal-related information throughout a sophisticated, multi-stage reconstruction pipeline to produce clearer images.
MGMAR's core innovation is a two-phase approach that combines a data-driven prior with iterative refinement. First, it generates a high-quality prior image by training a conditioned implicit neural representation (INR) using projections unaffected by metal. This INR is pretrained on paired corrupted and clean CT images, embedding prior knowledge directly into its parameters for robustness and faster convergence. An encoder analyzes the metal-corrupted scan along with a recursively constructed artifact map to capture global, metal-dependent artifact patterns.
In the second phase, this prior is integrated into a normalized MAR framework to complete the corrupted projections. Finally, a dedicated metal-conditioned correction network, where a metal mask modulates features via adaptive instance normalization, targets and suppresses residual secondary artifacts while meticulously preserving the patient's anatomical structures. This end-to-end, metal-aware design allows the system to adapt to the specific interference caused by different implants.
The results are compelling. On the public and challenging AAPM-MAR benchmark, a standard test in the field, MGMAR achieved state-of-the-art performance. It attained an impressive average final score of 0.89 across 29 diverse clinical test cases, demonstrating its effectiveness and potential for real-world clinical deployment to provide radiologists with more accurate, artifact-free diagnostic images.
- Uses a conditioned implicit neural representation (INR) pretrained on artifact-free/corrupted image pairs to create a robust prior.
- Integrates a metal-conditioned correction network with adaptive instance normalization to target residual artifacts while preserving anatomy.
- Achieved a state-of-the-art score of 0.89 on the AAPM-MAR benchmark, outperforming previous methods on 29 clinical cases.
Why It Matters
Enables more accurate diagnosis from CT scans for millions of patients with metallic implants, reducing uncertainty and improving surgical planning.