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

Background: Low-dose CT (LDCT) imaging reduces patient radiation exposure but introduces elevated noise levels that degrade image quality and undermine downstream clinical tasks such as diagnosis and quantitative analysis. Existing denoising approaches often require extensive diffusion steps that impede real-time clinical applicability. Purpose: To address this challenge, we propose a regularization-enhanced efficient diffusion probabilistic model (RE-EDPM), a rapid and high-fidelity denoising framework that incorporates residual guidance between low dose and full dose CT scans and employs hybrid perceptual and total variation regularization to preserve anatomical fidelity and diagnostic quality. Methods: RE-EDPM incorporates a residual-shifting mechanism into the forward diffusion process to better align the LDCT and FDCT distributions, followed by four reverse diffusion steps using a Swin-based U-Net backbone. A composite loss function combining pixel-level reconstruction, perceptual similarity (LPIPS), and spatially total variation (TV) is used to suppress spatially varying noise while preserving fine structural details. We evaluated RE-EDPM on a public LDCT benchmark dataset across different dose levels and anatomical sites. Quantitative performance was assessed using Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Visual Information Fidelity (VIFp), and RE-EDPM was compared against several state-of-the-art methods. Results: On public LDCT benchmarks at 10% dose for chest and 25% dose for abdomen, RE-EDPM achieved SSIM of 0.879 ± 0.068, PSNR of 31.60 ± 2.52 dB, and VIFp of 0.366 ± 0.121 for chest images, and SSIM of 0.971 ± 0.000, PSNR of 36.69 ± 2.54 dB, and VIFp of 0.510 ± 0.007 for abdominal images. Visualizations of residual and difference maps confirmed superior noise suppression and structural fidelity. Ablation studies and Wilcoxon signed-rank tests (p < 0.05) validated the significant contributions of the residual-shifting mechanism and each regularization component. RE-EDPM processes two 512 × 512 slices in approximately 0.25 s (≈ 0.125 s per slice) on modern high-performance hardware, supporting near-real-time clinical application. Conclusions: RE-EDPM offers robust LDCT denoising with minimal inference time, achieving an optimal balance between noise reduction and anatomical preservation. Its efficiency and high performance make it a strong candidate for real-time clinical deployment and for broader applications in transforming low-quality to high-quality medical imaging tasks.