Core Concepts
Addressing data consistency discrepancy in sparse-view CT reconstruction using a cascaded diffusion model with discrepancy mitigation.
Abstract
Sparse-view CT reconstruction poses challenges due to image degradation. The study introduces a Cascaded Diffusion with Discrepancy Mitigation (CDDM) framework to address this issue. By generating low-quality images in latent space and high-quality images in pixel space, the framework minimizes computational costs and mitigates the training-sampling gap induced by data consistency. Specialized Alternating Direction Method of Multipliers (ADMM) is employed for image gradients, enhancing regularization. Experimental results demonstrate CDDM's superior performance in high-quality image generation compared to existing methods.
Stats
Sparse-view Computed Tomography (CT) image reconstruction reduces radiation exposure.
Diffusion model-based approaches provide a potential solution but suffer from training-sampling discrepancy.
CDDM framework minimizes computational costs by moving inference steps from pixel space to latent space.
Specialized ADMM processes image gradients separately for better regularization.
Experimental results show CDDM outperforms existing methods in high-quality image generation.