Low-Dose CT Imaging Using a Regularization-Enhanced Efficient Diffusion Probabilistic Model

Abstract

Low-dose computed tomography (LDCT) reduces patient radiation exposure but introduces substantial noise that degrades image quality and hinders diagnostic accuracy. Existing denoising approaches often require many diffusion steps, limiting real-time applicability. We propose a Regularization-Enhanced Efficient Diffusion Probabilistic Model (RE-EDPM), a rapid and high-fidelity LDCT denoising framework that integrates a residual shifting mechanism to align low-dose and full-dose distributions and performs only four reverse diffusion steps using a Swin-based U-Net backbone. A composite loss combining pixel reconstruction, perceptual similarity (LPIPS), and total variation (TV) regularization effectively suppresses spatially varying noise while preserving anatomical structures. RE-EDPM was evaluated on a public LDCT benchmark across dose levels and anatomical sites. On 10 percent dose chest and 25 percent dose abdominal scans, it achieved SSIM = 0.879 (0.068), PSNR = 31.60 (2.52) dB, VIFp = 0.366 (0.121) for chest, and SSIM = 0.971 (0.000), PSNR = 36.69 (2.54) dB, VIFp = 0.510 (0.007) for abdomen. Visual and statistical analyses, including ablation and Wilcoxon signed-rank tests (p < 0.05), confirm significant contributions from residual shifting and regularization terms. RE-EDPM processes two 512x512 slices in about 0.25 s on modern GPUs, supporting near real-time clinical use. The proposed framework achieves an optimal balance between noise suppression and anatomical fidelity, offering an efficient solution for LDCT restoration and broader medical image enhancement tasks.