PT - JOURNAL ARTICLE AU - Yu, Boxiao AU - Dong, Yafei AU - Gong, Kuang TI - <strong>Whole-body PET Image Denoising Based on 3D Denoising Diffusion Probabilistic Model</strong> DP - 2024 Jun 01 TA - Journal of Nuclear Medicine PG - 242268--242268 VI - 65 IP - supplement 2 4099 - http://jnm.snmjournals.org/content/65/supplement_2/242268.short 4100 - http://jnm.snmjournals.org/content/65/supplement_2/242268.full SO - J Nucl Med2024 Jun 01; 65 AB - 242268 Introduction: Reducing PET dose is desirable to further lower radiation exposure to patients or enhance the hospital throughput. However, the degradation in image quality caused by lower counts severely limits its clinical applicability. The denoising diffusion probabilistic model (DDPM), an advanced distribution learning-based model, has shown outstanding performance in several medical image processing tasks. Currently DDPM is mainly investigated in 2D mode, which has limitations for whole-body PET image denoising given that PET is inherently a 3D imaging technique. In this work, we proposed and validated 3D DDPM for whole-body PET image denoising.Methods: For DDPM, noise was gradually injected into the images during the forward diffusion phase to perturb the input data. After network training, the reverse diffusion process was run to retrieve the desired noise-free data from noisy data samples. In this work, the patches of 96×96×96 randomly cropped from the whole-body PET data were used as the 3D DDPM’s network input, considering the available GPU memory. During the sampling stage, 3D data was first chunked along the axial direction, then fed sequentially into the 3D model, and finally combined to generate the final whole-body PET image. Fig. 1(A) shows the diagram of the proposed 3D DDPM.The Siemens Biograph Vision Quadra data from the Ultra-low Dose PET Imaging Challenge was utilized in our experiments, which comprised 377 whole-body 18F-FDG PET datasets. The 1/20 low-dose and normal-dose PET data were selected as low- and high-quality paired data for our model training and evaluation, which were randomly divided into training data (302 subjects), validation data (15 subjects), and testing data (60 subjects).The whole pipeline was implemented using Pytorch. The training batch size was set to 8, utilizing 8 NVIDIA A100 GPUs for distributed training. The training time was approximately 9 days and the testing time was an average of 28 minutes per test dataset. The 2D DDPM and 3D UNet-based image denoising were employed as reference methods.Results: Qualitative results shown in Fig. 1(B) indicated that the results produced by 3D UNet were overly smooth. In contrast, 2D DDPM and 3D DDPM generated more realistic denoising results. 3D DDPM significantly surpassed both 2D DDPM and UNet, revealing better structural details and more precise edge contours. The quantitative results shown in Fig. 2(A) matched with our qualitative observations. 3D DDPM achieved significantly better performance in both PSNR and SSIM than the other two methods. One advantage of DDPM is the ability to generate the uncertainty map. Fig. 2(B) shows the uncertainty map of the 3D DDPM results generated using 20 different random seeds, reflecting the model's high confidence in reproducibility.Conclusions: In this work, we proposed a 3D DDPM framework to enhance the quality of low-dose whole-body PET images. Our qualitative and quantitative results indicated the superiority of the proposed 3D DDPM over 2D DDPM and 3D UNet methods, demonstrating the advantage of extending diffusion models to 3D mode for whole-body PET image denoising.