Fast and detail-preserving low-dose CT denoising with diffusion model

Low-dose CT imaging is effective at reducing patient radiation exposure and prolonging equipment lifespan. However, it often leads to increased image noise and artifacts. Conventional convolutional neural networks may result in loss of detail and excessive smoothing. While diffusion models can mitigate this issue, their relatively low denoising efficiency limits their practical applicability. Recent studies focus on optimizing the sampling strategy. The denoising performance of diffusion models is heavily dependent on the UNet network, whereas our approach complements these studies by designing a more efficient noise prediction network to enhance denoising performance. First, we propose a novel and efficient extended perceptual downsampling (EEP), which broadens the perceptual field to capture richer information while maintaining original channel feature integrity. Notably, its computational effort and number of parameters are negligible. Second, we propose a symmetric-aware self-attention mechanism (SA self-attention) that leverages the symmetry and cross-scale similarity inherent in medical images to compute longitudinal and transverse interleaved attention maps at reduced scales, thereby lowering computational complexity. Third, a meticulous scanning protocol is implemented for the head and chest body models to acquire normal-dose, 10%, and 25% low-dose datasets. This dataset is made publicly available as part of this study. Comprehensive experimental results demonstrate that our method outperforms contemporary approaches in terms of performance and generalizability. Notably, our method achieves the highest performance in blind evaluations conducted by radiologists, with an inference time deemed clinically acceptable. • A Downsampling that enlarges the receptive field and preserves channel features. • A novel Self-attention that captures a broader range of dependencies. • A publicly available Low-dose CT dataset for head and chest regions is provided. • Our method can remove noise and exhibit robust effects at different sites and doses.