MuPET : Multiphysics and Noise Modelling Approach for Synthetic PET Data Generation – A Methodological Evaluation for 3D Printed Phantom Design

3D-printed phantoms offer customizability and cost-effective surrogates for research in medical imaging but their iterative design and testing often requires demand significant time and resources. In the context of positron emission tomography imaging, the design restrictions are furthermore due to use of radioactive tracers. Although simulations can be helpful in this optimizing the designs in early phases, most commonly used toolboxes are either computationally intensive or retrospective, capable of being performed only after initial prototyping. In this study we aim to introduce MuPET (Multiphysics PET), a computationally efficient framework to generate realistic synthetic PET data to accelerate the design, optimization, and evaluation of novel 3D-printed phantoms. The framework combines multiphysics based diffusion-reaction modelling to generate simulated activity maps followed by the integration of a scanner-specific pseudo-Poisson noise model and spatially variant Gaussian blurring to emulate noise characteristics. As a proof-of-concept, MuPET was validated against acquisition PET data acquired from a custom 3D-printed phantom and conventional synthetic lesion insertion method. The custom phantom consisted of 6 spherical targets (2 diameters – 15mm, 20mm x 3 target to background ratios – 2:1, 2.5:1, 3.33:1) and 3 non target 3D printed inserts. All the acquisitions were performed based on standard imaging guidelines. For the evaluation - volume averaged activity (in MBq), recovery coefficient (RC) and target contrast ratio (TCR), coefficient of variability (CoV) are utilized as metrics. Additionally, extracted activity profiles and ROI based histograms were compared for overlap. Wilcoxon rank sum test was utilized for statistical comparison of activity profiles with significance of p ∼ 0.05. Variability between volume averaged activity for computational geometry and physical phantom was less than 1.5%. RC _mean , TCR _max showed strong agreement with both acquisition and conventional simulation values falling within 10% bounds while RC _max showed slight underestimation. CoV values for MuPET (3.9-4.04%) closely matched acquired PET (3.75-3.91%). Activity profiles effectively captured the transition across various regions of the phantoms and the histogram jaccard overlap for targets was found to be 80% for 20mm and 60% for 15mm. No statistically significant differences were observed for the target activity profiles with p-values of ≥ 0.54 (vs acquisition) and ≥ 0.19 (vs conventional simulation). A notable achievement is MuPET’s computational efficiency, requiring approximately 10-15 minutes for a 35-minute simulation, alongside its flexibility in handling custom phantom designs. This faster, more adaptable pipeline aids in the design, optimization, and evaluation of novel 3D-printed phantoms.