Impact of Deep Learning-Based Time-of-Flight PET Images of Small Tumors Using a Human Anatomic Phantom

Time-of-flight (ToF) in PET improves image quality by enhancing the signal-to-noise ratio, and recent deep learning (DL)-based ToF (DL-ToF) methods further enhance tumor visibility and reduce noise. This study quantitatively investigates the effects of DL-ToF on PET images using the thoracoabdominal phantom simulating human anatomy. Methods: The phantom, containing optimized radioactivity of ¹⁸F-FDG in each organ and tumor, was scanned using a BGO crystal PET/CT machine. Imaging was performed at 6 acquisition times (1, 1.5, 2, 3, 5, and 10 min), with PET images reconstructed using the low, middle, and high levels of DL-ToF and non-ToF. The SUV_mean and SUV_max of each organ, lung, and liver tumors were measured for each acquisition time. Additionally, shape index maps were generated to assess pixel value changes and the impact of DL-ToF on image quality. Results: DL-ToF processing significantly improved tumor visibility and contrast, especially with the high-precision DL (HDL) model. For lung tumors, the [Formula: see text] increased from 3.72 (non-ToF) to 5.89 (HDL) at 10 min. Liver tumor [Formula: see text] also increased, with HDL yielding the highest [Formula: see text] (5.62). Shape index maps suggested that clearer tumor boundaries and enhanced contrast were obtained with high-precision DL-ToF. Clinical cases of lung and liver tumors demonstrated similar trends, with improved tumor delineation. Conclusion: DL-ToF can affect lesion visibility and image characteristics in a manner dependent on its processing level, underscoring the importance of understanding its behavior for clinical implementation.