An efficient deep unrolling network for sparse-view CT reconstruction via alternating optimization of dense-view...

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning

Background:

• Deep unrolling methods have advanced sparse-view computed tomography (SVCT) reconstruction by integrating model-based and deep learning techniques.
• Existing methods are computationally intensive, especially with large clinical datasets, limiting their practical application.
• There is a need for efficient SVCT reconstruction techniques that maintain high image quality while reducing computational demands.

Purpose of the Study:

• To develop a computationally efficient deep unrolling method for SVCT reconstruction.
• To maintain or improve the quality of reconstructed images from sparse-view projection data.
• To reduce the computational resources and runtime required for SVCT reconstruction, enhancing clinical applicability.

Main Methods:

• Decomposed the SVCT reconstruction into optimizing dense-view sinograms and images using proximal gradient methods.
• Employed deep neural networks for dense-view sinogram inpainting, image-residual learning, and image-refinement within an iterative framework.
• Focused on optimizing dense-view sinograms instead of full-view sinograms to reduce computational load and error propagation.

Main Results:

• Successfully reconstructed high-resolution images from real-size projection data with significantly reduced training parameters and fast inference time (0.09s/slice).
• Achieved superior quantitative and qualitative results compared to state-of-the-art methods, particularly in artifact suppression and detail preservation for low sparse ratios (1/12, 1/18).
• Demonstrated that dense-view sinogram inpainting accelerates computation, speeds up network convergence, and further enhances reconstruction quality.

Conclusions:

• The proposed dual-domain deep unrolling technique offers an efficient solution for SVCT reconstruction.
• The method achieves excellent reconstruction results with minimal computational resources, addressing a key limitation of previous approaches.
• This research paves the way for faster and more practical deep unrolling CT reconstruction methods for clinical data processing.