Deep Learning Strategies for PET and SPECT Image Enhancement

Abstract

Positron Emission Tomography (PET) and Single Photon Emission Computed Tomography (SPECT) are essential imaging modalities in nuclear medicine, providing critical diagnostic information for a wide range of medical applications, including oncology, cardiology, and neurology studies. However, the image quality of nuclear imaging could be negatively affected by various factors, including reduced radiation dose, limited spatial resolution, the partial volume effect, and the positron range effect. In this dissertation, a series of deep learning strategies are presented to address these challenges, with the aim of improving the safety, efficiency, and diagnostic accuracy of nuclear imaging.

In PET imaging, we addressed the challenges related to 3D PET image denoising and positron range effect. In Chapter 2, we developed DDPET-3D, a dose-aware diffusion-based framework for 3D PET imaging, which was validated on a large-scale, multi-institutional PET dataset. This framework demonstrated superior performance compared to baseline methods, with nuclear medicine physicists judging the denoised images to be visually comparable to full-dose images, highlighting its clinical potential. Additional evaluations of DDPET-3D, including multi-modal PET-CT denoising and PET image enhancement within a radiation therapy system, are provided in Appendix B. In Chapter 3, we addressed the noise and positron range effects in dynamic 82Rb cardiac PET imaging through a self-supervised noise-aware dynamic denoising and positron range correction framework. This approach enabled robust dynamic image denoising and positron range correction without paired training data. The presented results demonstrated the potential of improved image quantification, particularly in myocardial blood flow (MBF) assessments and other key clinical metrics. In the Appendix A, we introduced another PET image denoising method, the Unified Noise-aware Network (UNN), which employed a noise-aware mechanism to adaptively weight sub-denoising networks based on varying noise levels, providing robust and generalizable results.

In SPECT imaging, we introduced DiffSPECT-3D (Chapter 4), a diffusion-based framework for 3D cardiac SPECT image enhancement under various acquisition settings, including different low-dose and few-view settings. By enforcing projection/image data consistency and integrating anatomical priors, DiffSPECT-3D demonstrated superior generalizability across diverse acquisition settings without additional network re-training or fine-tuning. In Chapter 5, to mitigate the partial volume effect without the tedious manual segmentation steps, we proposed a segmentation-free partial volume correction method that improves quantitative precision while reducing radiation exposure to the patients by eliminating the need for additional contrast-enhanced CT or MRI scans. In Chapter 6, to further enhance image quality on stationary cardiac SPECT scanners, we developed a multi-angle reconstruction protocol to increase angular sampling. To preserve the stationary benefits of this scanner type, we introduced a deep-learning-based framework to synthesize multi-angle data from single-angle counterparts. In the Appendix D and Appendix E, we also explored image-domain and dual-domain transformer-based methods for few-view SPECT imaging. For all SPECT imaging projects, comprehensive validations using physical phantoms, animal studies, and human datasets were performed, demonstrating potential improvements in spatial resolution, diagnostic accuracy, and overall image quality.

The deep learning strategies proposed in this dissertation collectively enhance PET and SPECT imaging by addressing critical challenges such as dose reduction, noise suppression, and image resolution enhancement. These innovations contribute to safer, faster, and more accurate imaging, with the potential to improve diagnostic performance and patient outcomes, and establish a strong foundation for future advancements in AI-driven medical imaging technologies.