Unsupervised Tau PET Image Super-Resolution Using Cross-Domain Diffusion Posterior Sampling

Positron emission tomography images are often hindered by poor spatial resolution, impacting diagnostic accuracy and quantitation. In this paper, we propose an innovative unsupervised learning framework utilizing diffusion models to enhance the spatial resolution of PET images. Our method leverages a pre-trained diffusion model as a prior to inverse uniform degradation processes like Gaussian blur. By integrating cross-domain autoencoders, we enable the transfer of images between an unknown low-resolution dataset and a Gaussian blur dataset for effective posterior sampling. This framework was trained and validated on unpaired ${ }^{18} \mathrm{~F}$-Flortaucipir PET images from Siemens HR+ and GE Discovery MI scanners. We demonstrate that our approach reduces the Fréchet Inception Distance between super-resolved images and actual high-resolution images and improves the spatial resolution of PET images from legacy scanners. These results indicate that our method is promising for addressing super-resolution as well as general PET inverse problems, enhancing the utility of PET imaging in neurodegenerative disease research.