Semi-Supervised CT Denoising via Text-Guided Mamba Diffusion Models

Low-dose computed tomography (CT) reduces patient radiation exposure, yet it introduces noise and artifacts that can impair diagnostics. Recent supervised methods, particularly the diffusion method, have effectively minimized image smoothing while preserving details. However, iterative sampling incurs considerable time costs. In addition, obtaining a large number of paired low-dose and normal-dose CT images is clinically challenging, and imaging results vary across different body regions and devices. We propose a novel semi-supervised diffusion model. In addition, we design textual information related to different anatomical regions and devices in the denoising task and develop a text-guided approach that enables the model to generate accurate denoising results without requiring separate models for each region or device. An asymmetric encoder-decoder architecture has been implemented, featuring an encoder based on the state space model to enhance noise prediction capabilities. Furthermore, to increase sampling efficiency, a redraw feature has been proposed during the sampling process. It enables adjustment of the redraw rate according to the noise levels, ensuring detailed image preservation while significantly reducing time. We scanned head and chest phantoms using two devices, GE and Siemens, under two low-dose settings (10% and 25%), acquiring a total of 13 499 slices. The phantom dataset used for evaluation was scanned independently of the training set, including a total of 1400 images from two sites of two devices. Our approach is superior to the partially unsupervised and supervised methods by processing various body parts and device slices and is 100 times faster than denoising diffusion probabilistic models (DDPMs).