Sub-Millimeter 2D Intrinsic Resolution Recovery in Semi-Monolithic PET Detector using Neural Networks - a Monte Carlo Simulation

Accurate localization of the annihilation gamma-ray photon interaction points within detector crystals is critical for achieving high spatial resolution in positron emission tomography (PET). PET systems have achieved sub-millimeter resolution by employing pixelated detector blocks with small crystal size of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sim 1 \times 1 ~\text{mm}^{2}$</tex> in cross-section. While effective, this approach becomes expensive and technically complex when scaling to larger systems. Semi-monolithic crystal detectors offer a promising compromise by reducing readout complexity while maintaining high resolution through signal-based position estimation. In this study, we present an in-silico evaluation of two neural network architectures for estimating the monolithic and depth-of-interaction (DOI) coordinates in a semi-monolithic LYSO detector block, the building component of the Ultra-High Performance Brain PET (UHB-PET) system. The dataset was generated using GATE v8.0 simulations. We implemented a multi-layer perceptron (MLP) using 8 SiPM row-sum features (mimicking the realistic readout architecture), achieving resolutions of 0.45 and 0.59 mm in the monolithic and DOI directions, respectively. To improve accuracy near crystal edges, where MLP performance degrades, we developed an attention-augmented 1D CNN (A1D-CNN) using the full <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$8+8$</tex> summed SiPM signals projected along the monolithic and axial directions readout values. This model outputs probability distributions over spatial bins and computes estimates via expectation, achieving a 40 % reduction in the spatial variance of the prediction coordinates error. While the models perform well under ideal conditions, the lack of electronic noise and perfect ground truth limits generalizability. Future work will focus on evaluating the performance of this approach using experimental data, with the ultimate goal of achieving sub-millimeter resolution in UHB-PET system.