Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction for Sparse-View CT

A team of researchers led by Haodong Li has introduced CDPIR, a novel framework for Sparse-View CT (SVCT) reconstruction that addresses the critical challenge of out-of-distribution (OOD) artifacts. SVCT reduces radiation dose and improves temporal resolution by using fewer X-ray projections, but standard methods fail when scanner settings, protocols, or anatomy differ from training data. CDPIR tackles this by combining a Scalable Interpolant Transformer (SiT)—an evolution of the Diffusion Transformer (DiT)—with model-based iterative reconstruction. The SiT is trained across multiple CT datasets using Classifier-Free Guidance (CFG), randomly dropping conditioning to learn both domain-specific and domain-invariant priors, enabling generalization to unseen distributions.

During sampling, the transformer-based diffusion model exploits cross-distribution priors within a unified stochastic interpolant framework, allowing flexible control over multi-distribution-to-noise interpolation paths and decoupled sampling strategies. This enables CDPIR to adapt robustly to OOD reconstruction, alternating between data fidelity and diffusion updates to preserve fine details. Extensive experiments show CDPIR significantly outperforms existing SVCT methods, especially under OOD conditions, highlighting its potential for safer, lower-dose clinical imaging. The paper has been accepted by IEEE Transactions on Medical Imaging.