Diffusion Model-Based Reconstruction of Low-Dose PET to Standard-Dose PET

Abstract

Positron Emission Tomography (PET) is a powerful functional imaging modality widely used in clinical diagnostics and biomedical research, owing to its unique ability to visualize metabolic processes in vivo. However, due to inherent limitations in the image acquisition process, achieving high-quality PET images in clinical practices typically requires administering higher doses of radioactive tracers, which can introduce potential health risks to patients. Conversely, reducing the injected tracer dosage can mitigate these risks, but results in images that are significantly noisier and less clear, commonly referred to as low-dose PET (LPET) images. Given the critical role of PET in modern medicine, there is a growing research interest in developing advanced reconstruction techniques that can effectively reconstruct LPET images into high-quality standard-dose PET (SPET) images. The challenge of accurately and efficiently reconstructing high-quality PET images from low-dose, or even ultra-low-dose (UDPET), acquisitions is of paramount importance, as it holds the promise of enhancing diagnostic accuracy while minimizing patient exposure to radiation.

In this thesis, we begin by presenting an overview of the problem setting, detailing the inherent challenges and the motivations underpinning our research, thereby establishing a solid foundation for understanding the task at hand. Next, we provide a comprehensive literature review. Based on the insights drawn from this review, we propose a novel wavelet-informed diffusion WiD-PET method designed to address critical limitations of existing approaches—namely, low computational efficiency, suboptimal detail restoration, and spatial discontinuity. Building on this foundation, we further propose CWD-PET. Finally, we discuss potential research directions and outline opportunities for future refinement in the domain of low-dose PET (LPET) reconstruction using diffusion models.