Dual-Domain Cross-Prompt Learning for Efficient Sparse-View CT

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Sparse-view computed tomography (CT) reduces radiation dose but degrades image quality and diagnostic reliability.
• Deep unrolling networks offer a hybrid approach for CT reconstruction, combining optimization and data-driven methods.
• Existing methods struggle with simplistic priors, complex regularizers, and high computational costs, leading to over-smoothed images.

Purpose of the Study:

• To develop an advanced framework for sparse-view CT reconstruction that overcomes limitations of current deep unrolling networks.
• To improve image quality, artifact suppression, and preservation of fine structures in low-dose CT scans.
• To enhance computational efficiency and reduce reliance on complex regularizers and excessive iterations.

Main Methods:

• Proposed a Dual-domain Cross-Prompt Learning (DCPL) framework integrating prompt learning into unrolled gradient descent networks.
• Introduced an implicit pixel-wise learnable step size to adapt to image gradient heterogeneity.
• Incorporated learnable prompts into data-fidelity and regularization terms, with a cross-prompt mechanism for inter-domain interaction and stability.

Main Results:

• DCPL demonstrated consistent improvements in artifact suppression and fine structural preservation across multiple clinical benchmarks.
• The framework achieved robust reconstruction quality even under extremely sparse-view sampling conditions.
• Achieved significant reductions in parameters, higher inference efficiency, and required fewer unrolled iterations compared to existing methods.

Conclusions:

• The DCPL framework effectively addresses limitations in sparse-view CT reconstruction, offering superior performance.
• This approach enables high-quality image reconstruction with reduced radiation exposure and improved diagnostic reliability.
• DCPL presents a computationally efficient and robust solution for advanced medical imaging applications.