Sparse-view sampling for computed tomography (CT) reduces radiation exposure in medical imaging but suffers from severe artifacts due to ill-posed nature of under-sampled reconstruction. While diffusion posterior sampling (DPS) leverages data-driven priors for improved results, its stochastic nature impedes clinical reproducibility, and inefficient noise-perturbed likelihood estimation limits practicality. This work proposes the probabilistic ordinary differential equation (ODE)-driven Deterministic Diffusion Posterior Sampling (ODDPS) framework to enable stable, high-quality sparse-view CT reconstruction by combining physical constraints with deterministic generative sampling. ODDPS reformulates stochastic diffusion into a probability flow ODE, ensuring deterministic sampling from noise to image. In this work, a dual-domain solver is proposed, which integrates diffusion processes in the image domain and sinogram domain, and decouples the problem into a physicsbased likelihood term and a diffusion-based prior term under the Bayesian framework. Specifically, the prior term is solved using the recently proposed first-order solver of diffusion probabilistic model, while the likelihood term adopts an efficient physicsbased likelihood gradient to avoid neural backpropagation in each iteration. This design bridges tomographic physics and the learned diffusion prior within an unified framework, ensuring the consistency between the reconstruction results and the physical constraints of CT imaging. Evaluated on public datasets, ODDPS achieved average improvements of 1.19 dB and 5% in peak signalto-noise ratio and structural similarity index measured under the 30 views, and is 48.6% faster reconstruction than DPS. Ablation studies confirmed near-zero variance in outputs and significant artifact suppression. By unifying ODE-driven deterministic sampling with dual-domain diffusion, ODDPS delivers reproducible, high-fidelity sparse-view CT reconstructions. Its efficiency and Repeatability advance the clinical translation of diffusion models for inverse problems.
Shen et al. (Tue,) studied this question.