Advancing Medical Imaging: The CDPIR Framework for Sparse-View CT

Modern medical imaging faces a constant trade-off between image quality and patient safety. Specifically, Sparse-View CT reconstruction offers a promising solution by reducing radiation exposure and improving temporal resolution. However, reducing the number of projections often results in significant artifacts and performance degradation when scanners or anatomical variations differ from the training data. To address these out-of-distribution (OOD) challenges, researchers have developed the Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction (CDPIR) framework.

This innovative approach integrates cross-distribution diffusion priors with model-based iterative reconstruction methods. The system utilizes a Scalable Interpolant Transformer (SiT) backbone, which effectively extends the Diffusion Transformer architecture. Consequently, the model establishes a unified stochastic interpolant framework that leverages Classifier-Free Guidance (CFG). By randomly dropping conditioning during training, the model learns transferable priors. These priors encourage domain-invariant anatomical structures while allowing for specific appearance modulation. Therefore, the system maintains accuracy even across diverse clinical environments.

Clinical Benefits of Enhanced Sparse-View CT Reconstruction

The CDPIR framework significantly improves adaptation to OOD reconstruction by exploiting globally sensitive transformer-based diffusion models. Furthermore, the framework alternates between data fidelity and sampling updates during the reconstruction process. This strategy ensures superior detail preservation compared to existing deep learning methods. Consequently, clinicians can expect more reliable diagnostic images even when using lower radiation doses or unconventional scanning protocols. In addition, the robustness of this model highlights its potential clinical value in challenging imaging scenarios across various hospital settings.

Frequently Asked Questions

What is Sparse-View CT reconstruction?

Sparse-View CT reconstruction is a technique that uses fewer X-ray projections to create a cross-sectional image. This method significantly reduces the total radiation dose to the patient and speeds up the acquisition process, although it traditionally risks lower image quality without advanced algorithms.

How does the CDPIR framework handle out-of-distribution data?

The CDPIR framework uses a Scalable Interpolant Transformer to learn cross-distribution priors. By leveraging Classifier-Free Guidance, the model differentiates between universal anatomical structures and scanner-specific details. This allows it to maintain high performance across different machines and patient populations.

Disclaimer: This content is for informational and educational purposes only. It does not constitute medical advice or professional services. Always seek the advice of a physician or other qualified health provider regarding a medical condition. Refer to the latest local and national guidelines for clinical practice.

References

Li H et al. Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction for Sparse-View CT. IEEE Trans Med Imaging. 2026 Apr 23. doi: 10.1109/TMI.2026.3687173. PMID: 42024952.