Multi-Material Decomposition Using Spectral Diffusion Posterior Sampling.

Area of Science:

• Medical Imaging
• Computational Imaging
• Materials Science

Background:

• Accurate material decomposition is essential for spectral computed tomography (CT) applications.
• Current methods often face challenges with accuracy, stability, and computational efficiency.

Purpose of the Study:

• Introduce Spectral Diffusion Posterior Sampling (Spectral DPS), a novel framework for one-step reconstruction and multi-material decomposition in spectral CT.
• To evaluate the performance of Spectral DPS against existing methods using simulations and physical phantom studies.

Main Methods:

• Spectral DPS integrates advanced prior information from unconditional network training with an analytical physical system model.
• The framework builds upon the general DPS approach for nonlinear inverse problems, incorporating techniques like jumpstart sampling and Jacobian approximation.
• Evaluations were conducted on simulated dual-layer and kV-switching spectral systems, as well as a physical cone-beam CT (CBCT) test bench.

Main Results:

• Spectral DPS demonstrated significant reductions in sampling variability and computational costs compared to Baseline DPS.
• The method outperformed the Conditional Denoising Diffusion Probabilistic Model (DDPM) in imaging accuracy and robustness across diverse imaging protocols.
• In physical phantom studies, Spectral DPS achieved a 1% error in mean density estimation and avoided artifacts seen in other methods.

Conclusions:

• Spectral DPS provides a superior approach for accurate, stable, and fast material decomposition in spectral CT.
• This general framework effectively combines learning-based priors with physics-based spectral models, applicable to various spectral CT systems and basis materials.