Prompting Lipschitz-Constrained Network for Multiple-in-One Sparse-View CT Reconstruction.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Deep learning for sparse-view computed tomography (SVCT) faces challenges with provable Lipschitz constraints and high storage costs for multi-view models.
• Current methods struggle to ensure theoretical convergence and practical applicability across diverse sparse sampling settings.

Purpose of the Study:

• To develop a novel deep learning framework for multiple-in-one SVCT reconstruction that addresses limitations in provable constraints and storage efficiency.
• To enable a single model to handle various sparse view configurations effectively.

Main Methods:

• Introduced LipNet, an explicitly provable Lipschitz-constrained network, ensuring theoretical convergence.
• Developed PromptCT, a storage-saving deep unfolding framework embedding LipNet for multiple-in-one SVCT.
• Integrated an explicit prompt module for discriminative knowledge of different sparse sampling settings.

Main Results:

• PromptCT demonstrated superior performance over benchmark algorithms in both simulated and real data experiments.
• Achieved higher-quality SVCT reconstructions with significantly lower storage costs.
• Theoretically proved LipNet's boundary property and Lipschitz continuity, confirming algorithm convergence.

Conclusions:

• PromptCT offers an effective and efficient solution for multiple-in-one SVCT reconstruction.
• The framework provides a practical approach for clinical applications by reducing storage requirements and ensuring reconstruction quality.
• The explicit Lipschitz constraints and prompt module enhance the reliability and versatility of deep learning-based SVCT.