Abstract

Positron emission tomography (PET) provides functional information that is essential for disease diagnosis, treatment assessment, and biomedical research. However, the development and validation of PET imaging methods often require simulated imaging data with known ground truth, as well as large clinical datasets that are realistic and scalable. These requirements are difficult to satisfy using clinical studies alone, motivating the development of virtual PET imaging techniques. Physics-based simulation can model PET acquisition with high accuracy, but its computational cost limits its large-scale usage. Meanwhile, conventional digital phantom-based image generation approaches often lack the anatomical diversity and heterogeneous tracer uptake patterns observed in clinical PET images. To address these limitations, this dissertation develops advanced computational pipelines for virtual PET imaging, progressing from efficient PET image simulation to deep learning-based PET image synthesis. The first study developed FAST-PET as a fast and accurate analytical PET simulation method implemented using Siemens reconstruction software e7tools. FAST-PET models PET acquisition through precise forward projection, scatter estimation, and random estimation matched to scanner geometry and statistics while avoiding the computational burden of full Monte Carlo simulation. Validation with a NEMA image quality phantom and clinical PET simulations demonstrated that FAST-PET reproduced quantitative image characteristics with high accuracy. By reducing computation time by approximately 100-fold, FAST-PET provides an accurate and efficient platform for large-scale PET simulation, quantitative imaging research, and virtual imaging trials. The second study extended this direction by developing a noise-aware system generative model, NASGM, for deep learning-based PET simulation. NASGM was designed to reduce dependence on proprietary reconstruction software while enabling efficient simulation of PET images. By conditioning image generation on acquisition duration and using a dual-domain discriminator in both spatial and frequency domains, NASGM captured count-dependent image characteristics and preserved quantitative image properties. Across multiple evaluations, NASGM-generated images showed strong agreement with reference PET images in quantitative uptake, noise behavior, and texture features. The generated images also supported clinically relevant tumor detection tasks and were visually difficult to distinguish from reference PET images in observer studies, demonstrating the potential of deep learning-based simulation for scalable PET imaging research. The third study moved beyond simulation of acquisition conditions toward anatomy-conditioned clinical PET image synthesis. A pretrained domain-adapted diffusion model, PAD, was developed to generate heterogeneous PET images from uniform organ activity maps. PAD uses a natural-image pretrained diffusion decoder to improve training stability and data efficiency, with a domain adapter incorporated to transform the decoder output from the natural-image domain to the PET image domain. A two-phase coarse-to-fine training strategy is further used to support realistic full-resolution PET synthesis. PAD-generated images preserved organ-level quantitative uptake, with concordance correlation coefficients exceeding 0.92 for mean SUV measurements of major organs. The synthesized images also reproduced realistic noise and radiomic characteristics and achieved tumor segmentation performance comparable to real PET images. Finally, PAD’s successful application to XCAT-derived activity maps further demonstrated its compatibility with digital phantom-based anatomical priors. In conclusion, this dissertation advances virtual PET imaging across multiple levels of realism and control. FAST-PET provides an efficient analytical simulation foundation, NASGM introduces a flexible deep learning-based strategy for acquisition-duration-aware PET simulation, and PAD enables realistic clinical PET image synthesis from simplified anatomical activity inputs. These methods expand the capabilities of virtual PET imaging by enabling realistic and controllable PET data generation at scale, supporting data augmentation, protocol optimization, virtual imaging trials, and the development and validation of downstream PET image processing methods.

Committee Chair

Kooresh Shoghi

Committee Members

Daniel Thorek; Jingqin Luo; Joseph O'Sullivan; Yuan-Chuan Tai

Degree

Doctor of Philosophy (PhD)

Author's Department

Interdisciplinary Programs

Author's School

McKelvey School of Engineering

Document Type

Dissertation

Date of Award

7-23-2026

Language

English (en)

Download

Share

COinS