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

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Sparse-View CT (SVCT) offers reduced radiation dose and improved temporal resolution.
• Clinical application of SVCT is limited by artifacts from view reduction and domain shifts.
• Out-of-distribution (OOD) data variations lead to significant performance degradation in existing SVCT methods.

Purpose of the Study:

• To develop a robust reconstruction framework for SVCT that addresses artifacts and OOD data challenges.
• To improve the clinical utility of SVCT by enhancing image quality and stability across diverse data distributions.
• To introduce a novel method that leverages cross-distribution diffusion priors for improved SVCT reconstruction.

Main Methods:

• Proposed a Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction (CDPIR) framework.
• Integrated cross-distribution diffusion priors from a Scalable Interpolant Transformer (SiT) with iterative reconstruction.
• Utilized Classifier-Free Guidance (CFG) across multiple datasets and a unified stochastic interpolant framework.

Main Results:

• CDPIR demonstrated superior detail preservation and artifact reduction in SVCT reconstructions.
• The framework significantly outperformed existing methods, especially under OOD conditions.
• Achieved state-of-the-art performance by balancing data fidelity and sampling updates.

Conclusions:

• CDPIR offers a robust solution for SVCT reconstruction, particularly in challenging OOD scenarios.
• The proposed method shows significant potential for clinical value in diverse imaging environments.
• The developed framework enhances the reliability and applicability of low-dose CT imaging.