Multi-stage dual-domain progressive network with synergistic training for sparse-view CT reconstruction

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiology

Background:

• Sparse-view computed tomography (SVCT) reduces radiation exposure and scan time.
• SVCT image reconstruction faces artifacts, limiting diagnostic value.
• Current methods lack flexibility for varied sparse-view conditions.

Purpose of the Study:

• To develop a flexible and effective reconstruction method for diverse sparse-view CT scenarios.
• To address the limitations of existing SVCT reconstruction techniques.
• To improve image quality and diagnostic utility in low-projection CT.

Main Methods:

• Proposed a Multi-view Synergistic Training Strategy (MSTS) for a single model adaptable to various sparse-view settings.
• Introduced the Multi-stage Dual-domain Progressive Reconstruction Network (MDPRNet) for progressive sinogram-to-image reconstruction.
• Incorporated a Cross-stage Feature Adapter (CFA) within MDPRNet for enhanced spatial detail.

Main Results:

• MDPRNet demonstrated superior performance across a wide range of sparse-view scenarios, including ultra-sparse cases.
• The method successfully generated images with fine spatial details, overcoming common artifacts.
• Experiments on clinical and custom datasets validated MDPRNet's effectiveness against state-of-the-art methods.

Conclusions:

• MDPRNet offers a robust and adaptable solution for sparse-view CT reconstruction.
• The proposed MSTS and MDPRNet architecture significantly improve image quality and diagnostic potential in low-dose CT.
• This work provides a flexible deep learning framework for diverse SVCT applications.