CT-free Attenuation Correction of Whole-body PET Images via Direct Diffusion Bridge

Attenuation correction of PET images is a crucial step in obtaining accurate quantification in PET/CT and PET/MRI studies. Typically, structural images such as CT images are required to calculate attenuation coefficients and correct for the effects of gamma-ray attenuation. However, CT-based attenuation correction methods pose a concern due to CT artifacts and radiation exposure. Recently, there have been attempts to synthesize CT from MRI or non-attenuation corrected PET (NAC PET) using generative models, such as Generative Adversarial Networks (GAN). However, utilizing the synthetic CT for generating attenuation-corrected PET (AC PET) has resulted in significant errors in whole-body PET scans, thus limiting its clinical applications. In this study, we have adopted the recently proposed diffusion bridge model (DBM) [1] to address this issue, which learns the transition trajectory between two different distributions. We aimed to generate the attenuation map ($\mu$-map) from NAC PET and subsequently perform the PET attenuation correction using the $\mu$-map. We conducted a comparative analysis of $\mu$-maps generated using U-NET, conditional Generative Adversarial Networks (cGAN), and Diffusion Bridge Models (DBM). The evaluation metrics utilized were normalized mean squared error (NMSE), peak signal-to-noise ratio (PSNR), and Structural Similarity Index Measure (SSIM). DBM exhibited superior performance relative to the existing methods. Following this, we conducted attenuation correction of 18F-FDG PET using attenuation maps produced by each method. CT-based AC PET served as the ground truth for this comparison, using NMSE and PSNR as the assessment metrics. DBM achieved significant enhancements in performance across all metrics when compared to traditional approaches