Abstract
Metal artifacts in Computed Tomography (CT) scans, caused by high-density implants, significantly degrade image quality and clinical utility. Traditional supervised deep learning approaches for Metal Artifact Reduction (MAR) are limited by the lack of paired clinical data—scans of the same patient with and without metal. This paper presents a self-supervised, cross-domain learning framework that trains exclusively on metal-free data. By integrating sinogram completion with a novel differentiable reconstruction loss, the proposed method outperforms existing supervised models and preserves fine anatomical details more effectively than standard approaches.
Introduction
When a patient has a metal implant, such as a hip prosthesis or dental filling, it absorbs X-rays much more strongly than surrounding bone or tissue. In the raw data collected by the CT scanner—the sinogram—these metal objects create "shadows" or missing regions. Upon reconstruction into a 3D image, these gaps manifest as bright streaks and dark bands known as metal artifacts. These artifacts can obscure critical anatomical features, hide tumors, and impede accurate radiation dose calculations for cancer treatment.
While deep learning has shown promise in MAR, most models rely on supervised learning with paired datasets. Since obtaining identical clinical scans with and without metal is physically impossible, researchers typically rely on synthetic data. However, models trained on "fake" artifacts often struggle to generalize to real-world clinical scans. This work addresses these challenges by introducing a self-supervised approach that learns from metal-free data.
Methods
The proposed framework bypasses the need for paired data by treating MAR as a self-supervised inpainting problem. Instead of learning from "bad-to-good" pairs, the model learns by intentionally "breaking" clean data and attempting to restore it.
The architecture consists of several key components:
- SinoNet (Sinogram Completion): The network takes a clean, metal-free sinogram and hides a portion of it to simulate a "metal trace." It is then trained to fill in these missing pixels.
- Cross-domain Learning & FBP Loss: To ensure the sinogram repairs are physically consistent, the authors introduce a differentiable Filtered Back Projection (FBP) layer. This allows the network to evaluate how its sinogram repairs look as a reconstructed image during training. The FBP reconstruction loss penalizes repairs that result in image-domain streaks.
- ImgNet (Refinement): A secondary module operates in the image domain to further smooth lingering secondary artifacts.
- Metal Trace Replacement: To maintain clinical fidelity, the final image is not purely a CNN output. Instead, the original corrupted sinogram is used, with only the pixels within the metal trace replaced by the network's predictions. This ensures the uncorrupted parts of the scan remain 100% authentic.
Results
The model was evaluated using the DeepLesion dataset (1,000 CT slices with 90 simulated metal masks), the MICCAI 2015 Head and Neck dataset for generalization, and real-world clinical data from patients with bilateral hip prostheses.
Quantitative analysis on the DeepLesion dataset demonstrated the superiority of the proposed method. It achieved a Root Mean Square Error (RMSE) of 34.20 HU, significantly outperforming standard Linear Interpolation (50.31 HU) and even supervised state-of-the-art models like DuDoNet (38.00 HU). Qualitatively, the method effectively removed challenging artifacts, such as the "dark bridge" between bilateral hip implants, which often persists in other algorithms. Furthermore, the model generalized well to head CTs despite being trained on torso scans, indicating it learned general principles of X-ray physics rather than specific anatomy.
Conclusion
This self-supervised framework offers a robust solution for MAR by leveraging the underlying geometry of healthy sinograms and incorporating differentiable physics directly into the learning process. By prioritizing clinical fidelity through metal trace replacement, it avoids the "hallucinations" common in other AI models. Future work will focus on extending the framework to 3D volumes to account for cross-slice dependencies and improving the accessibility of raw sinogram data in clinical settings.
Summary updated