Physics-Guided Dual-Domain Plug-and-Play ADMM for Low-Dose CT Reconstruction

TL;DR

This paper introduces a physics-guided, plug-and-play iterative reconstruction method for ultra-low-dose CT that leverages a self-supervised deep denoiser, achieving high-quality images at significantly reduced radiation doses.

Contribution

It presents a novel dual-domain PnP framework with a self-supervised denoiser trained via a 2-stage Noise-to-Noise scheme, enhancing ultra-low-dose CT reconstruction quality.

Findings

Achieves 70-80% dose reduction while maintaining diagnostic quality.

Outperforms existing deep learning and supervised PnP methods in texture and detail preservation.

Effectively reduces artifacts and preserves subtle tissue contrast.

Abstract

Ultra-low-dose CT (ULDCT) imaging can greatly reduce patient radiation exposure, but the resulting scans suffer from severe structured and random noise that degrades image quality. To address this challenge, we propose a novel Plug-and-Play model-based iterative reconstruction framework (PnP-MBIR) that integrates a deep convolutional denoiser trained in a 2-stage self-supervised Noise-to-Noise (N2N) scheme. The method alternates between enforcing sinogram-domain data fidelity and applying the learned image-domain denoiser within an optimization, enabling artifact suppression while maintaining anatomical structure. The 2-stage protocol enables fully self-supervised training from noisy data, followed by high-dose fine-tuning, ensuring the denoiser's robustness in the ultra-low-dose regime. Our method enables high-quality reconstructions at $\sim$70--80\% lower dose levels, while maintaining diagnostic fidelity comparable to standard full-dose scans. Quantitative evaluations using Gray-Level Co-occurrence Matrix (GLCM) features -- including contrast, homogeneity, entropy, and correlation -- confirm that the proposed method yields superior texture consistency and detail preservation over standalone deep learning and supervised PnP baselines. Qualitative and quantitative results on both simulated and clinical datasets demonstrate that our framework effectively reduces streaks and structured artifacts while preserving subtle tissue contrast, making it a promising tool for ULDCT reconstruction.