A total variation prior unrolling approach for computed tomography reconstruction

Area of Science:

• Medical Imaging
• Deep Learning
• Image Reconstruction

Background:

• Deep learning enhances computed tomography (CT) reconstruction.
• Current methods often replace prior terms with neural networks instead of unrolling them.
• Unrolling prior terms could improve CT reconstruction performance.

Purpose of the Study:

• Propose a primal-dual network (PD-Net) by unrolling both data fidelity and total variation (TV) prior terms.
• Preserve image edges and textures in CT reconstructions.
• Develop a specialized unrolling approach for improved CT reconstruction.

Main Methods:

• Derived a Chambolle-Pock (CP) algorithm instance for CT reconstruction.
• Discovered TV prior updates images via divergences in each iteration.
• Applied CNNs to yield feature map divergences, ensuring they match TV prior roles.

Main Results:

• Assessed PD-Net using Low-Dose CT and Piglet datasets.
• PD-Net effectively preserved structural and textural information compared to ground truth.
• Demonstrated superior performance over conventional CT reconstruction methods.

Conclusions:

• PD-Net framework is feasible for CT reconstruction.
• The study inspires direct unrolling of hand-crafted prior terms to CNNs.
• Promising results indicate potential for further development in unrolling approaches.