Iterative reconstruction for limited-angle CT using implicit neural representation.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Limited-angle computed tomography (CT) is inherently ill-posed, leading to artifacts with traditional reconstruction methods.
• Supervised deep learning approaches for limited-angle CT face challenges with generalization and the need for large paired datasets.
• Existing methods struggle to reconstruct high-fidelity images from sparse projection data.

Purpose of the Study:

• To develop an iterative neural reconstruction framework for limited-angle CT that overcomes data dependency.
• To improve the accuracy and stability of neural network-based CT reconstruction.
• To enable high-fidelity image reconstruction from limited projection data without extensive training datasets.

Main Methods:

• Proposed an iterative neural reconstruction framework utilizing a coordinate-based neural representation.
• Formulated tomographic reconstruction as a convex optimization problem solved by a deep neural network.
• Employed a differentiable projection layer for network optimization and introduced prior-based weight initialization for stability.

Main Results:

• The proposed self-supervised method significantly outperforms existing iterative and learning-based approaches.
• Demonstrated effective restoration of anatomical features and structures on XCAT and Mayo Clinic datasets.
• Quantitative evaluations (NRMSE, PSNR, SSIM) and visual inspections confirmed superior image quality compared to methods starting from scratch.

Conclusions:

• The developed framework successfully reconstructs high-fidelity CT images from limited-angle x-ray projections.
• The data-free, self-supervised approach with informed initialization enhances medical image reconstruction.
• This methodology holds significant potential for various clinical applications in medical imaging.