Anatomically and Metabolically Informed Deep Learning Low-Count PET Image Denoising

Recent progress in deep learning (DL) positron emission tomography (PET) image denoising has shown great potential for generating high-count images from low-count data, thus minimizing radiation exposure during imaging. However, DL models typically over-smooth images with details blurred, leading to quantification underestimation. In this study, we propose leveraging both organ and lesion priors obtained by auxiliary segmentation tools to make denoised images more anatomically and metabolically credible. Specifically, the DL training process is adjusted by semantic-weighted guidance. This approach forces class-wise alignment of standard uptake values and histogram distributions, and adaptively assigns weights to each semantic class. Experiments were performed on a dataset with 200 cases acquired by a Siemens Biograph mCT PET/CT system. Results showed that incorporating semantic information improved image quality and reduced the normalized root-mean-square error (NRMSE) on liver, lung, and lesion by 7.06%, 5.58%, and 1.21% on average, respectively, compared to the DL baseline without segmentation guidance.