Beyond 0.29±0.02 mm intrinsic spatial resolution based on monolithic crystals using convolutional neural network: a simulation study

High spatial resolution is important for a Positron Emission Tomography (PET) system. Monolithic crystal based detectors are promising to deliver sub-millimeter intrinsic spatial resolution, due to their high sensitivity, continuous position encoding, and cost-effectiveness. However, their full potential is hindered by two key obstacles: the non-linear photoelectric response function that is used to deduce the annihilation event position, and the nonoptimal coupling structures between the crystal and photodetectors. To address these challenges, we propose several solutions, targeting less than 0.3 mm intrinsic spatial resolution. First, we develop a deep learning based positioning algorithm that can more accurately pin down annihilation event positions from the non-linear light distribution, significantly enhancing intrinsic spatial resolution. Second, we propose a crystal dimension optimization strategy to mitigate the detrimental "edge effect". Finally, we establish a quantitative evaluation mechanism, leveraging statistical comparison, to systematically identify the optimal crystal-sensor coupling structure. Extensive simulation-based validation demonstrates that our integrated approach achieves an exceptional average intrinsic spatial resolution of 0.29±0.02mm, with a peak local resolution of 0.2mm. This unprecedented resolution opens new frontiers for PET imaging, enabling demanding applications such as single cell tracking and dynamic imaging of the rodent brain.