ESTRO 2025 - Abstract Book

S2998

Physics - Image acquisition and processing

ESTRO 2025

denoising and reconstruction tasks. However, previous work has shown that generating high-resolution images using diffusion models directly is still diffcult. The main goal of this study is to propose an LDCT denoising method based on a diffusion model, which improves the image denoising effect by combining global and local feature information, thereby avoiding the computational bottleneck and insufficient accuracy of the diffusion model when directly generating high-resolution images. Material/Methods: This study designed a dual-branch denoising diffusion model architecture, in which the global branch downsamples the image to obtain global features. In contrast, the local branch captures local detail information by inputting image blocks. This global-local feature fusion method integrates high-level features through the self-attention mechanism so that the model can retain the global texture consistency of the image during denoising and reduce the texture and contrast differences in the edge areas of the image blocks. In addition, the model uses a multi-scale feature fusion module to fuse global and local information, thereby effectively improving the accuracy of denoising and reconstructed images.

Results: The experimental results show that the proposed diffusion model outperforms the traditional method and the denoising model based on convolutional neural network (CNN) in LDCT denoising performance, can effectively eliminate image noise and artifacts, and retain image details to avoid over-smoothing. The comparative experimental results show that this method presents higher texture clarity and contrast in the denoised image. The quantitative evaluation results also show that the proposed model is superior to other denoising methods in terms of mean absolute error (MAE=9.66±1.3HU), peak signal-to-noise ratio (PSNR=43.08±1.3dB), and structural similarity index (SSIM=0.983±0.02), and has significant robustness and accuracy.