GPU-based list-mode TOF PET image reconstruction with complete correction techniques

Positron emission tomography (PET) is essential for the early diagnosis of cancer, neurological disorders, and cardiovascular diseases. However, achieving high-quality PET images remains challenging due to the complex physical factors involved and the trade-off between reconstruction accuracy and computational efficiency.

This study aimed to develop QuanTOF, a GPU-accelerated PET reconstruction framework integrating comprehensive physical corrections and advanced modeling to enhance image quality while maintaining clinical practicality.

QuanTOF employs a GPU-accelerated Bayesian penalized-likelihood reconstruction algorithm with time-of-flight (TOF) and point spread function (PSF) modeling. It incorporates full corrections for attenuation, normalization, random coincidences, and scatter coincidences. A memory-efficient TOF single scatter simulation (SSS) algorithm enabled on-the-fly scatter correction without storing full TOF sinograms. Validation included Monte Carlo simulations, clinical phantom experiments, and blinded reader studies using a Siemens Biograph Vision PET/CT scanner.

In phantom studies, QuanTOF achieved uniformity comparable to commercial systems and resolved 2.4 mm hot rods in Derenzo phantoms, with peak-to-valley ratios of 2.61 (3.2 mm rods) and 1.29 (2.4 mm rods). Scatter correction time was reduced to $0.05\pm 28.87$ s, two orders of magnitude faster than existing methods. Clinicians rated QuanTOF images significantly higher (average score: 3.90 vs. 2.15 for clinical images) in reader studies. Reconstruction times remained clinically acceptable (e.g., 2.78 s for NEMA phantom scatter correction) CONCLUSIONS: QuanTOF balances accuracy and efficiency through GPU-optimized physics modeling and memory-efficient algorithms. It delivers high-resolution PET images with diagnostic confidence, demonstrating potential for clinical oncology, neurology, and cardiology applications.