Optimizing a Parameterized Plug-and-Play ADMM for Iterative Low-Dose CT Reconstruction.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Reducing X-ray radiation exposure in CT scans is crucial for patient safety.
• Model-based iterative reconstruction (MBIR) algorithms are used for low-dose CT (LdCT) but require manual parameter tuning and prior knowledge.
• Existing MBIR methods face challenges in balancing image quality with reduced radiation dose.

Purpose of the Study:

• To develop a deep learning (DL)-based strategy for low-dose CT (LdCT) image reconstruction.
• To simultaneously address prior knowledge design and parameter selection within a unified optimization framework.
• To enhance the robustness and performance of LdCT reconstruction algorithms.

Main Methods:

• A parameterized plug-and-play alternating direction method of multipliers (3pADMM) was proposed for penalized weighted least-squares models.
• A deep learning approach was integrated to optimize the parameterized plug-and-play (3p) prior and its parameters concurrently.
• The framework was trained and validated using extensive clinical patient datasets.

Main Results:

• The proposed 3pADMM framework successfully optimized the 3p prior and related parameters simultaneously.
• Experimental results demonstrated significant improvements in noise-induced artifact suppression compared to existing algorithms.
• Enhanced edge detail preservation was observed in LdCT images reconstructed with the proposed method.

Conclusions:

• The DL-based 3pADMM strategy offers a robust and effective solution for LdCT image reconstruction.
• This method overcomes limitations of traditional MBIR by automating prior knowledge integration and parameter optimization.
• The approach shows promising clinical applicability for improving LdCT imaging quality and patient safety.