RDDM: One-step CT denoising achieves 15ms inference with state-of-the-art fidelity

Researchers from Rensselaer Polytechnic Institute (Jianxu Wang, Qing Lyu, Ge Wang) have introduced RDDM (Residual-Driven Drifting Model), a novel approach to low-dose CT (LDCT) denoising that dramatically accelerates inference while improving image fidelity. The model addresses a key limitation of diffusion models: multi-step iterative inference that makes real-time use impractical. RDDM incorporates a residual drifting field formed by attractive and repulsive forces between LDCT and normal-dose CT residuals, enabling single-step denoising without sacrificing quality.

The team developed three variants (RDDM-Base, RDDM-Fine, RDDM-Coarse) to balance detail preservation and noise suppression. RDDM-Fine achieves state-of-the-art results across supervised baselines, with a PSNR of 41.2 dB, SSIM of 0.987, and an outstanding FID of 5.87 — indicating highly realistic anatomical textures. Critically, inference takes only about 15 milliseconds per 512×512 slice, making it viable for real-time clinical workflows. The model outperforms popular diffusion-based methods like DDPM while requiring orders of magnitude less computation.