| Abstract: |
Attenuation correction (AC) is a critical challenge in preclinical PET/MRI imaging due to high-resolution requirements and limited photon counts, which exacerbate attenuation effects and reconstruction errors. Conventional methods, including CT and MR-based AC, provide standard solutions but are limited by additional radiation exposure, artifacts, and reliance on accurate segmentation. Deep learning (DL) approaches have emerged as promising alternatives, yet they often produce over-smoothed images and fail to preserve fine structural details. In this study, we present a conditional denoising probabilistic model (DDPM) to generate high-quality attenuation corrected PET (PETAC) images directly from non-attenuation corrected PET (PETNAC) inputs. The model was trained and validated on a curated dataset of micro-phantoms and FDG PET rat scans acquired with a Bruker 7T preclinical PET/MRI system. We evaluated multiple loss function strategies, including MSE, SSIM and VGG perceptual losses, individually and in combination. The combined MSE+SSIM+VGG loss achieved the best results, improving quantitative metrics such as PSNR, SSIM, RMSE, SNR, CNR, and SUV accuracy while preserving anatomical details. Our findings demonstrate that diffusion-based AC offers a robust, high-fidelity alternative to conventional approaches in preclinical PET/MRI imaging. |