Diffusion transformer model with compact prior for low-dose PET reconstruction

Physics in Medicine and Biology, Journal Year: 2025, Volume and Issue: 70(4), P. 045015 - 045015

Published: Jan. 20, 2025

Abstract Objective. Positron emission tomography (PET) is an advanced medical imaging technique that plays a crucial role in non-invasive clinical diagnosis. However, while reducing radiation exposure through low-dose PET scans beneficial for patient safety, it often results insufficient statistical data. This scarcity of data poses significant challenges accurately reconstructing high-quality images, which are essential reliable diagnostic outcomes. Approach. In this research, we propose diffusion transformer model (DTM) guided by joint compact prior to enhance the reconstruction quality imaging. light current research findings, present pioneering integrates and models optimization. combines powerful distribution mapping abilities with capacity transformers capture long-range dependencies, offering advantages reconstruction. Additionally, incorporation lesion refining block alternating direction method multipliers recovery capability regions preserves detail information, solving blurring problems areas texture details most deep learning frameworks. Main . Experimental validate effectiveness DTM image quality. achieves state-of-the-art performance across various metrics, including PSNR, SSIM, NRMSE, CR, COV, demonstrating its ability reduce noise preserving critical such as structure texture. Compared baseline methods, delivers best denoising preservation levels, 10%, 25%, 50%, even ultra-low-dose level 1%. shows robust generalization on phantom datasets, highlighting adaptability varying conditions. Significance approach reduces ensuring early disease detection decision-making, promising tool both applications.

Language: Английский

Multimodal feature‐guided diffusion model for low‐count PET image denoising

Medical Physics, Journal Year: 2025, Volume and Issue: unknown

Published: March 18, 2025

Abstract Background To minimize radiation exposure while obtaining high‐quality Positron Emission Tomography (PET) images, various methods have been developed to derive standard‐count PET (SPET) images from low‐count (LPET) images. Although deep learning enhanced LPET they rarely utilize the rich complementary information MR Even when are used, these typically employ early, intermediate, or late fusion strategies merge features different CNN streams, failing fully exploit properties of multimodal fusion. Purpose In this study, we introduce a novel feature‐guided diffusion model, termed MFG‐Diff, designed for denoising with full utilization MRI. Methods MFG‐Diff replaces random Gaussian noise and introduces degradation operator simulate physical processes imaging. Besides, it uses cross‐modal guided restoration network modality‐specific provided by utilizes feature module employing cross‐attention mechanisms positional encoding at multiple levels better Results Under four counts (2.5%, 5.0%, 10%, 25%), generated our proposed showed superior performance compared those produced other networks in both qualitative quantitative evaluations, as well statistical analysis. particular, peak‐signal‐to‐noise ratio improved more than 20% under 2.5% count, structural similarity index 16%, root mean square error reduced nearly 50%. On hand, had significant correlation (Pearson coefficient, 0.9924), consistency, excellent evaluation results SPET Conclusions The method outperformed existing state‐of‐the‐art models can be used generate highly correlated consistent obtained

Language: Английский

Whole-body PET image denoising for reduced acquisition time

Frontiers in Medicine, Journal Year: 2024, Volume and Issue: 11

Published: Sept. 30, 2024

Purpose A reduced acquisition time positively impacts the patient's comfort and PET scanner's throughput. AI methods may allow for reducing without sacrificing image quality. The study aims to compare various neural networks find best models denoising. Methods Our experiments consider 212 studies (56,908 images) 7MBq/kg injected activity evaluate using 2D (RMSE, SSIM) 3D (SUVpeak SUVmax error regions of interest) metrics. We tested 2.5D ResNet, Unet, SwinIR, MedNeXt, UX-Net. have also compared supervised with unsupervised CycleGAN approach. Results conclusion model denoising is MedNeXt. It improved SSIM on 38.2% RMSE 28.1% in 30-s 16.9% 11.4% 60-s when original 90-s at same discrepancy dispersion.

Language: Английский