Limited-angle computed tomography with deep image and physics priors.

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Computed tomography (CT) is a vital 3D imaging technique used across diverse scientific fields.
• Limited angle reconstruction is a significant challenge in CT, leading to image artifacts and hindering quantitative analysis.
• Existing methods struggle with distortions caused by incomplete projection data, often due to sample constraints.

Purpose of the Study:

• To develop an advanced method for accurate 3D image reconstruction from limited angle CT data.
• To address the ill-posed nature of limited angle reconstruction by integrating generative models with physics-based approaches.
• To improve the quantitative evaluation of reconstructed images in challenging CT scenarios.

Main Methods:

• A self-training generative model was employed to learn the mapping from partial projections to the 3D object.
• The approach combines data likelihood and image prior terms within a single, end-to-end trained deep network.
• Total-variation regularization and an alternating direction method of multipliers solver were used to manage noise and optimize reconstruction.

Main Results:

• The proposed method effectively constrains the solution for limited angle reconstruction problems.
• Numerical results demonstrate significant improvements in image quality across various missing angle ranges and noise levels.
• The integrated deep network approach proved computationally tractable and enhanced reconstruction performance.

Conclusions:

• The developed generative model-based approach offers a robust solution for limited angle computed tomography reconstruction.
• This technique successfully mitigates artifacts and improves the accuracy of 3D structural analysis from incomplete data.
• The findings highlight the potential of deep learning and physics-based methods for advancing CT imaging applications.