Mitigating Data Consistency Induced Discrepancy in Cascaded Diffusion Models for Sparse-view CT Reconstruction

TL;DR

This work tackles the ill-posed nature of sparse-view CT by introducing CDDM, a cascaded diffusion framework that performs low-quality image generation in a latent space and high-quality refinement in pixel space. A one-step reconstruction combines data consistency via a specialized ADMM with a discrepancy mitigation diffusion path to align the learned prior with the data manifold. Empirical results on Walnut and AAPM datasets show that CDDM improves PSNR and maintains SSIM relative to baselines while reducing computational burden through latent-space inference, offering a practical diffusion-based solution for extreme undersampling in CT and a template for similar inverse problems.

Abstract

Sparse-view Computed Tomography (CT) image reconstruction is a promising approach to reduce radiation exposure, but it inevitably leads to image degradation. Although diffusion model-based approaches are computationally expensive and suffer from the training-sampling discrepancy, they provide a potential solution to the problem. This study introduces a novel Cascaded Diffusion with Discrepancy Mitigation (CDDM) framework, including the low-quality image generation in latent space and the high-quality image generation in pixel space which contains data consistency and discrepancy mitigation in a one-step reconstruction process. The cascaded framework minimizes computational costs by moving some inference steps from pixel space to latent space. The discrepancy mitigation technique addresses the training-sampling gap induced by data consistency, ensuring the data distribution is close to the original manifold. A specialized Alternating Direction Method of Multipliers (ADMM) is employed to process image gradients in separate directions, offering a more targeted approach to regularization. Experimental results across two datasets demonstrate CDDM's superior performance in high-quality image generation with clearer boundaries compared to existing methods, highlighting the framework's computational efficiency.

Figures (6)

• Figure 1: Overview of CDDM framework for sparse-view CT reconstruction. A standard diffusion process in latent space generates low-quality CT images, with the image condition provided by specialized ADMM. Noise is then added to the images in pixel space and the diffusion process is utilized again to generate high-quality CT images. Specialized ADMM treats the gradients of different directions separately, and the discrepancy mitigation uses another diffusion process to correct errors induced by data consistency.
• Figure 2: Representative results of 8-view CT reconstruction performance on Walnut dataset. From left to right, ground truth, ADMM, DDS, and our proposed CDDM. The three columns show reconstruction images of the axial, coronal, and sagittal plane from top to bottom. For ease of presentation, the rectangular coronal and sagittal images were center cropped to become squares.
• Figure 3: Representative results of 8-view CT reconstruction performance on AAPM dataset. From left to right, ground truth, DDS, and our proposed CDDM. The three columns show reconstruction images of the axial, coronal, and sagittal plane from top to bottom.
• Figure 4: The relationship between noise strength $t_0$ added to the latent diffusion results and the final reconstruction images.
• Figure 5: The relationship between total inference steps $T$ in the pixel diffusion process and the final reconstruction images.