Changes in FDG-PET Images of Small Lung and Liver Masses Caused by the Deep Learning-based Time-of-flight Processing

Abstract

Purpose: The deep learning time-of-flight (DL-ToF) aims to replicate the ToF effects through post-processing, applying deep learning-based enhancement to PET images. This study evaluates the effectiveness of DL-ToF using a chest-abdomen phantom that simulates human anatomical structures. Methods: The 3 DL-ToF intensities (Low-DL-ToF: LDL, Middle-DL-ToF: MDL, High-DL-ToF: HDL) were adopted for the PET image of the chest-abdomen phantom. We assessed the mean SUV of the liver, kidneys, and soft tissue, as well as the maximum SUV of lung and liver tumors. Additionally, non-ToF images were subjected to 3 types of filtering. Texture analysis and shape index maps were used to evaluate filter effects. Results: No significant differences were observed in the mean SUV between the 3 DL-ToF and non-ToF images. LDL sharpened lung tumors and smoothed liver tumors, while HDL exhibited more pronounced sharpening effects. Conclusion: The DL-ToF produces image effects similar to ToF in PET imaging.