Multi-scale adaptive residual cold diffusion model for Low-Dose CT denoising

Low-Dose Computed Tomography (LDCT) images are typically severely affected by noise and artifacts, and it is crucial to enhance CT image quality while preserving diagnostic accuracy. In recent years, with the extension of classical diffusion models to Cold diffusion, the fast generation of high-quality images has been achieved through deterministic sampling capabilities. This paper proposes a Multi-scale Adaptive Residual Cold Diffusion Model (MarCoDiff) for LDCT image denoising tasks. Firstly, MarCoDiff employs a cosine mean-preserving degradation (CMPD) operator to effectively simulate the physical degradation process of CT images and accelerate sampling steps. Additionally, a cosine scheduling strategy is integrated to enhance learning at intermediate time steps, and improve the stability of reverse process sampling. Secondly, MarCoDiff constructs a multi-scale adaptive residual denoising network. In this network, we develop an adaptive frequency adjustment module (AFAM) to adaptively adjust the fusion of different frequency components between input features and time step embedding features, effectively preventing over-smoothing or noise in denoised images. We propose a multi-scale residual adaptive block (MRAB) to reconstruct features at each scale from coarse to fine, adapt to texture variations, restore intricate details and edges, and dynamically eliminate anisotropic noise. We design a hybrid-dilated-convolution context block (HCB) that employs mixed symmetric dilation rates to form dilated convolution chains, capturing multi-scale contextual information without further resolution reduction. The model implements a two-stage denoising strategy to efficiently generate higher-quality denoised CT images by alleviating noise and artifact effects. Extensive comparative experiments are conducted on three LDCT datasets. Experimental results demonstrate that the proposed method outperforms other approaches while exhibiting enhanced generalization capability under varying noise levels. We have made our code publicly available at https://github.com/levvvis/ MarCoDiff.