Abstract
Metal artifact correction is a challenging problem in cone beam computed tomography (CBCT) scanning. Metal implants inserted into the anatomy cause severe artifacts in reconstructed images. Widely used inpainting-based metal artifact reduction (MAR) methods require segmentation of metal traces in the projections as a first step, which is a challenging task. One approach is to use a deep learning method to segment metals in the projections. However, the success of deep learning methods is limited by the availability of realistic training data. It is laborious and time consuming to get reliable ground truth annotations due to unclear implant boundaries and large numbers of projections. We propose to use X-ray simulations to generate synthetic metal segmentation training dataset from clinical CBCT scans. We compare the effect of simulations with different numbers of photons and also compare several training strategies to augment the available data. We compare our model’s performance on real clinical scans with conventional region growing threshold-based MAR, moving metal artifact reduction method, and a recent deep learning method. We show that simulations with relatively small number of photons are suitable for the metal segmentation task and that training the deep learning model with full size and cropped projections together improves the robustness of the model. We show substantial improvement in the image quality affected by severe motion, voxel size under-sampling, and out-of-FOV metals. Our method can be easily integrated into the existing projection-based MAR pipeline to get improved image quality. This method can provide a novel paradigm to accurately segment metals in CBCT projections.
The Challenge of Metal Artifacts in CBCT
When a patient with a metal implant—such as a dental screw or an orthopedic plate—undergoes a Cone Beam Computed Tomography (CBCT) scan, the high-density metal absorbs X-rays so strongly that it creates "metal artifacts." These appear as streaks, shadows, and "blooming" effects in the final 3D reconstruction, often obscuring critical anatomical details and hindering accurate diagnosis.
To mitigate these effects, clinicians employ Metal Artifact Reduction (MAR) algorithms. Most projection-based MAR methods follow a three-step process: identifying the "metal trace" in the raw 2D X-ray projections, removing that data, and "inpainting" (interpolating) the missing areas. However, the first step—accurately segmenting metal in 2D projections—is notoriously difficult. Manual labeling is impractical for large datasets, and simple thresholding often fails due to noise, motion blur, and the presence of metal outside the 3D reconstruction volume (out-of-field-of-view or out-of-FOV).
A Deep Learning Approach to Segmentation
The researchers proposed a robust deep learning approach using a modified U-Net architecture to automate 2D metal segmentation. To overcome the lack of labeled clinical data, they developed a sophisticated synthetic data generation pipeline.
The pipeline involves taking existing clinical scans, segmenting the metal in 3D, and using Monte Carlo simulations to project that metal back into 2D. This creates pairs of "metal-corrupted" projections and corresponding "ground truth" metal masks. The model was trained on 10 clinical scans covering various extremity anatomies (knee, wrist, ankle, etc.), resulting in over 9,000 image crops for training. A hybrid strategy using both full-size projections and random crops was employed to help the network learn both global anatomical context and fine-grained metal boundaries.
Key Insights: The Power of Noisy Data
A pivotal finding of this study is that "noisy" data is actually better for training. Instead of computationally expensive "clean" simulations, the researchers intentionally used low photon counts (300–1400 photons per pixel) in their Monte Carlo simulations.
This introduced realistic noise that acted as a natural regularizer, making the model significantly more robust when applied to real-world clinical hardware. Furthermore, because the U-Net operates directly in the 2D projection domain, it can identify metal that is out-of-FOV in the 3D volume—a major advantage over traditional 3D-to-2D projection methods. The model also showed remarkable robustness to patient motion, learning the visual characteristics of motion-blurred metal that typically causes 3D-based methods to fail.
Clinical Impact and Results
The proposed "Noisy (full+crops)" model achieved a Dice Similarity Coefficient (DSC) of 94.8% on clinical benchmarks, outperforming conventional region-growing (90.5%) and consistency-check methods (91.0%). Crucially, the hybrid training strategy reduced the False Positive Rate in metal-free scans to nearly zero, ensuring the algorithm does not erroneously remove non-metal structures.
Qualitatively, the method demonstrated substantial improvements in image quality for scans affected by severe motion and out-of-FOV metal, preserving bone details and reducing streaking artifacts. While currently focused on extremity imaging, this approach provides a novel paradigm for accurate metal segmentation that can be easily integrated into existing MAR pipelines. Future work will explore expanding the dataset to include more diverse clinical scenarios, such as dental CBCT.
Summary updated