CT-SDM: A Sampling Diffusion Model for Sparse-View CT Reconstruction across All Sampling Rates

TL;DR

This work tackles the challenge of reconstructing high-quality SVCT images across arbitrary sampling rates without training separate models for each rate. It introduces CT-SDM, a diffusion-based framework that operates in the sinogram domain, where the forward degradation mirrors projection sampling and inference progressively restores full-view sinograms from undersampled data. A grouped-random sampling strategy enhances training diversity, while TACoS-based inference stabilizes reconstruction, enabling robust performance across rates. Empirical results on LDCT and TCGA-KIRC datasets show superior accuracy and stability compared with rate-specific baselines, highlighting the method's practical potential for flexible, dose-aware CT imaging.

Abstract

Sparse views X-ray computed tomography has emerged as a contemporary technique to mitigate radiation dose. Because of the reduced number of projection views, traditional reconstruction methods can lead to severe artifacts. Recently, research studies utilizing deep learning methods has made promising progress in removing artifacts for Sparse-View Computed Tomography (SVCT). However, given the limitations on the generalization capability of deep learning models, current methods usually train models on fixed sampling rates, affecting the usability and flexibility of model deployment in real clinical settings. To address this issue, our study proposes a adaptive reconstruction method to achieve high-performance SVCT reconstruction at any sampling rate. Specifically, we design a novel imaging degradation operator in the proposed sampling diffusion model for SVCT (CT-SDM) to simulate the projection process in the sinogram domain. Thus, the CT-SDM can gradually add projection views to highly undersampled measurements to generalize the full-view sinograms. By choosing an appropriate starting point in diffusion inference, the proposed model can recover the full-view sinograms from any sampling rate with only one trained model. Experiments on several datasets have verified the effectiveness and robustness of our approach, demonstrating its superiority in reconstructing high-quality images from sparse-view CT scans across various sampling rates.

Figures (9)

• Figure 1: (a) DL based SVCT reconstructions are face performance drop when train and test on different sampling rates (i.e, Sampling Mismatch). (b) The proposed method the proposed method replace the image degradation in forward process of diffusion model as the projection view sampling. Choosing a certain sampling rate as a start point at inference, the proposed model can reconstruct SVCT images under any sampling rates.
• Figure 2: The overall framework of the proposed methods. The forward process of Sampling Diffusion Model is to determine the undersampled CT measurements (i.e., sinograms) at each sampling step $t$. And a sampling step $t$ is corresponded to a specific sampling rate $\alpha_{t}$. The reverse process is aims to recover the full-view sinogram $y_{0}$ from the measurements $y_{T}$ obtained at a certain start sampling rate $\alpha_{T}$, utilizing the networks trained for undersampled data recovering. Both the forward and reverse process diffusion are designed in sinogram domain to allow the simulation of accoutre data sampling.
• Figure 3: The Group-Random Sampling Schedule. The grouped-random sampling method divides all sampling angles into $c$ orderly and equally spaced groups to ensure comprehensive and partially random angle coverage.
• Figure 4: Performance comparison of various methods at different sampling rates on the LDCT dataset. Each method is trained under 60 views(red line), and the results are evaluated on test data with varying numbers of projections: 116, 100, 74, 60, 55, 40, 30, and 23.
• Figure 5: Box plots showing the performance of various methods at different sampling rates on the LDCT dataset. Each method is trained under 60 projections, and the results are evaluated on test data with varying numbers of projections: 116, 100, 74, 60, 55, 40, 30, and 23.