Boosting positron emission tomography reconstruction with positional encoding-based deep image prior

In this paper, we leverage the structured foundation of deep image prior to delve into the complexities of positron emission tomography (PET) image reconstruction. We aim to underscore the potential of deep learning in overcoming inherent challenges associated with PET imaging. Acknowledging the limitations of conventional supervised learning in this domain, we propose an innovative unsupervised approach employing deep neural networks to enhance PET reconstruction. A central focus of our study revolves around the spectral bias issue that arises during PET image reconstruction. To tackle this challenge, we introduce a comprehensive framework that incorporates Gaussian Fourier features and Uniform Positional encoding. Our approaches undergo rigorous testing on both Brainweb data and naive rat data, revealing a noticeable improvement in image reconstruction performance. This underscores the efficacy of our framework in advancing PET imaging methodologies.This article is part of the theme issue 'Frontiers of applied inverse problems in science and engineering'.