Integrating compartment models and neural networks towards clinically reliable PET tracer dynamics prediction

Dynamic PET imaging enables the characterization of tracer kinetics over time, providing richer quantitative information with respect to conventional static acquisitions. The recent development of Long Axial Field-of-View (LAFOV) scanners, such as the uExplorer, has made high-resolution dynamic imaging technically feasible. However, standard dynamic protocols can last up to 60 minutes, limiting patient throughput and hindering clinical adoption due to their cost and duration. This thesis investigates whether shorter dynamic acquisitions could be made clinically feasible by extrapolating the late portion of the time-activity curve (TAC) from early data from total-body acquisitions obtained with the uExplorer scanner. The study focused on liver scans and sought to estimate the unknown 20–60 minute TAC segment from the first 20 minutes data. Three methods were tested on a dataset of 19 patients: a two-tissue compartment model (2TCM), a Neural Ordinary Differential Equation (NODE) based model and a hybrid model, conceptually combining the first two approaches. Results show that NODE obtained the worst performances, while the hybrid model generally produced more accurate predictions than 2TCM, especially at the 60-minute point, crucial for accurate clinical reporting. Furthermore, at 60 minutes, when the hybrid model outperformed 2TCM, the improvements were up to 3.4 times greater than the maximum advantage achieved by 2TCM over the hybrid, highlighting hybrid model potential. These results demonstrate the hybrid model effectiveness over NODE and 2TCM, establishing a new lower bound on achievable performance. They also pave the way for future studies to further develop this promising approach, ultimately contributing to the optimization of acquisition protocols towards the clinical implementation of dynamic PET imaging.