Answers are grounded in this paper and cite their sources.
Demonstrates improved image quality in low-dose CT scans, suggesting enhanced diagnostic utility.
Reducing patient radiation exposure in computed tomography (CT) imaging necessitates the use of low-dose CT (LDCT) scans. However, LDCT scans are accompanied by increased quantum noise and various artifacts that degrade image quality. Traditional denoising methods often lead to oversmoothing, failing to preserve crucial texture details. To address these issues, we utilize a diffusion model approach, specifically employing a pre-trained Score-based Generative Model (SGM), for the denoising of LDCT images. We employ the Variance Preserving (VP) SDE framework and train our SGM on normal-dose CT (NDCT) data. Our novel approach bypasses the initial stages of the SGM inference process by directly utilizing LDCT images as if they were generated in these early steps. However, directly inserting LDCT images into the inference process is not effective due to differences in noise characteristics between LDCT noise and Gaussian noise. To mitigate this issue, we introduce the Langevin Markov Chain Monte Carlo (LMC) technique, which refines the LDCT images, aligning them more closely with the expected noise distribution encountered during training. Our method outperformed conventional approaches like FBPConvNet in preserving fine textures and reducing noise, as demonstrated by improved LPIPS scores. The application of the LMC technique further improved performance across all metrics, including RMSE, SSIM, and LPIPS, compared to the approach without LMC. In this work, we leverages the generative capabilities of SGM, providing a significant advancement in LDCT image processing without the need for paired training data.
No takes yet. Share an insight, caveat, or question.
Jung et al. (2025) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: