Deep Learning-Based Synthesis Algorithms for Positron Emission Tomography (PET) Imaging

Abstract

Medical imaging plays a vital role in healthcare by enabling non-invasive visualization of anatomical and physiological processes. Among modalities, Positron Emission Tomography (PET) provides metabolic insights crucial for diagnosing cancers such as lung cancer and lymphoma. However, PET’s reliance on radioactive tracers poses radiation risks, particularly during sequential scans used to monitor treatment response. While low-dose PET reduces exposure, it compromises image quality and diagnostic accuracy.

This thesis addresses the challenge of synthesizing high-quality PET images from low-dose scans, known as low-to-high PET synthesis. Existing methods enhance image quality but often lack generalizability across dose levels and neglect the longitudinal nature of clinical imaging. Moreover, current CNN-based methods capture local features well but struggle with global dependencies essential for structural consistency.

To overcome these limitations, this thesis proposes three novel deep learning-based PET synthesis approaches:

SS-AEGAN introduces a self-supervised, adaptive residual estimation GAN to improve generalizability across varying low-dose levels, achieving robust synthesis performance on public datasets.

Trans-synGAN leverages baseline PET/CT to guide follow-up PET synthesis, explicitly modeling spatial and metabolic transformations across time points to address sequential scan inconsistencies.

Hybrid-CMLP integrates CNNs for local detail extraction with MLPs for global structure modeling, offering improved fidelity and contextual coherence with low computational overhead.

Validated on three datasets—including real low-dose PET—the proposed methods consistently outperform state-of-the-art techniques in both image quality and generalizability, providing more reliable, radiation-efficient molecular imaging.