Generative Super-Resolution PET Imaging with Fourier Diffusion Models

Abstract

Neurological Positron Emission Tomography (PET) is a critical imaging modality for diagnosing and studying neurodegenerative diseases like Alzheimer's disease. However, the inherent low spatial resolution of PET images poses significant challenges in clinical settings. This work introduces a novel Generative Super-Resolution (GSR) approach using Fourier Diffusion Models (FDMs) to enhance the spatial resolution of PET images. Unlike traditional methods, FDMs leverage the time-dependent Modulation Transfer Function (MTF) and Noise Power Spectrum (NPS) to generate high-resolution, low-noise images from low-resolution inputs. Our method was evaluated using simulated data derived from High-Resolution Research Tomograph (HRRT) PET images with 2 mm resolution. The results demonstrate that FDMs significantly outperform existing techniques, including conditional diffusion models and image-to-image Schr\"odinger bridge, across several metrics, including structural similarity and noise suppression. Our simulation results highlight the potential of FDMs to generate high-quality 2mm resolution reconstructions given 4mm resolution input PET data.