Impact of Re-framing High-Dose Scans for the Training of Neural Networks for Denoising Low-Dose PET

In PET, patient radiation exposure is mainly determined by the administered amount of radiotracer and its metabolism. By lowering the activity, the radiation dose is minimized at the cost of increased statistical fluctuations in the data; the latter increases the image noise level. Many artificial intelligence (AI) approaches have been proposed for image denoising of low-dose PET scans. Supervised AI algorithms require the use of labeled datasets, with the high-dose PET image as the target and the low-dose image as the input. This approach assumes that the performance of the PET systems is the same in the two activity regimes. However, the contribution of random coincidences, pile-up and dead-time effects depend on event rates. This study focuses on comparing two approaches to generating the training pair and their impact on the performance of a 3D U-Net using Monte Carlo simulations. The package GATE, v.9.1, was used to simulate a standard BMI XCAT phantom and a Biograph Vision Quadra PET/CT. Digital twins of 5-min
¹⁸
F-FDG-PET scans for an activity of 3 MBq/kg were obtained to simulate two low-dose scenarios: injected low-dose, and injected high-dose re-framed to low-dose. The results showed that both methods of generating low-dose datasets yield comparable results without degradation in image quality. However, differences were observed in the random fraction for the high-dose dataset, i.e., the texture of the data noise changed. This study shows that using high-dose data is suitable for training of AI-based denoising algorithms on image space; nevertheless, denoising algorithms working on the projection space might overestimate the noise level within the data. Future work will include other activity concentrations and scan durations, incorporating motion, and the use of other radioisotopes.