Diffusion prior-guided implicit neural representation for metal artifact reduction in sparse-view CT reconstruction.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medical Imaging

Background:

• Sparse-view computed tomography (SVCT) reduces radiation dose but causes artifacts, especially with metal implants.
• Existing methods for SVCT and metal artifact reduction (MAR) are often separate or require extensive paired data.
• Joint SVCT and MAR (SVMAR) is challenging due to artifact complexity and data limitations.

Purpose of the Study:

• To develop a self-supervised framework for joint sparse-view computed tomography and metal artifact reduction (SVMAR).
• To overcome limitations of existing supervised methods requiring large, difficult-to-obtain clinical datasets.
• To leverage denoising diffusion probabilistic models (DDPM) and implicit neural representation (INR) for improved SVMAR.

Main Methods:

• An INR is optimized for initial reconstruction using sparse-view data outside metal traces.
• A self-supervised framework alternates between MAR (inpainting metal-trace regions) and SVCT (refining INR with diffusion priors and corrected sinograms).
• The process utilizes diffusion priors and Poisson blending for artifact correction and data fidelity.

Main Results:

• High-quality image reconstruction achieved on simulation and clinical datasets without large paired datasets.
• The proposed method outperformed existing techniques like IndudoNet+ on out-of-distribution data.
• Significant improvements in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) were observed across various view settings.

Conclusions:

• The self-supervised INR framework effectively addresses the SVMAR problem.
• The method offers a practical and generalizable solution for real-world scenarios lacking ground-truth images.
• This approach advances low-dose CT imaging by enabling artifact reduction without extensive supervised training data.