GPDM: generation-prior diffusion model for accelerated direct attenuation and scatter correction of whole-body18F-FDG...

Objective.Accurate attenuation and scatter correction (ASC) is essential in positron emission tomography (PET) for reliable visual interpretation and quantitative analysis. Conventional correction methods based on computed tomography or magnetic resonance imaging have limitations related to accuracy, radiation exposure, and practical applicability. Although deep neural networks, particularly generative adversarial network (GAN)-based approaches, have shown promise for generating attenuation- and scatter-corrected PET (ASC PET) images from non-attenuation and non-scatter-corrected PET (NASC PET) images, their performance is still limited by instability during training and mode collapse. This study proposes a novel framework to generate high-quality ASC PET images from NASC PET images.Approach.We developed a generation-prior diffusion model (GPDM) for ASC PET image generation. The proposed framework is based on the denoising diffusion probabilistic model, but instead of initiating sampling from a completely unrelated image distribution, it starts from a distribution similar to the target ASC PET images. This prior distribution, termed thegeneration-prior(GP), guides the sampling process toward the desired image domain. By leveraging this prior information, GPDM reduces the number of required sampling steps and improves the refinement of the generated ASC PET images.Main results.Experimental results demonstrated that GPDM outperformed existing methods for generating ASC PET images from NASC PET images. The proposed framework achieved higher image generation accuracy while substantially reducing sampling time. These results indicate that incorporating a GP into the diffusion process improves both efficiency and image quality compared with conventional approaches.Significance.The proposed GPDM framework addresses key limitations of conventional GAN-based and diffusion-based methods for ASC in PET imaging. By enabling more efficient and accurate generation of ASC PET images from NASC PET images, GPDM has the potential to provide a robust alternative to conventional correction strategies and to advance practical PET imaging workflows.