Unsupervised CT Metal Artifact Reduction by Plugging Diffusion Priors in Dual Domains

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Metallic implants in computed tomography (CT) generate artifacts, hindering accurate diagnosis.
• Supervised deep learning methods for metal artifact reduction (MAR) require challenging-to-obtain paired simulated data, limiting clinical applicability.
• Existing unsupervised MAR methods often operate solely in either the image or sinogram domain.

Purpose of the Study:

• To develop an unsupervised metal artifact reduction (MAR) method using diffusion models.
• To overcome the limitations of supervised methods by eliminating the need for paired training data.
• To improve the performance of unsupervised MAR by leveraging dual-domain processing.

Main Methods:

• A diffusion model is trained on artifact-free CT images.
• Diffusion priors are iteratively applied in both the sinogram and image domains to restore metal artifact-affected regions.
• Temporally dynamic weight masks are designed for image-domain fusion, enabling dual-domain processing.

Main Results:

• The proposed unsupervised dual-domain diffusion model significantly outperforms existing unsupervised MAR methods.
• The method achieves superior visual results compared to both supervised and unsupervised methods on clinical datasets.
• Effectiveness is validated both qualitatively and quantitatively on synthetic datasets.

Conclusions:

• The dual-domain diffusion model offers a powerful unsupervised approach for metal artifact reduction (MAR) in CT imaging.
• This method enhances diagnostic accuracy by effectively removing metal artifacts without requiring paired training data.
• The approach shows significant potential for clinical application in CT imaging with metallic implants.