|
内容記述 |
In light-sharing positron emission tomography (PET) detectors, Anger logic enables identification of crystals smaller than the photodetector size. This approach, which employs a 2D position histogram and pixel-to-crystal mapping, is widely adopted due to its hardware implementation simplicity. In contrast, independent readout systems allow interaction detection using light distribution analysis via photodetector arrays, as each photodetector collects signals independently. Both approaches require detector-specific recalibration during manufacturing, maintenance, and replacement. In this work, we develop a PET scanner-wide crystal identifier using convolutional neural networks (CNNs). Using independent photodetector signals obtained experimentally, the CNN processes crystal addresses derived via Anger logic as training data. When applied to a small animal PET scanner with 126 light-sharing PET detectors, the model is trained using 5 million events from only 40 detectors after denoising, which particularly includes removal of inter-crystal scattering (ICS). By learning light distribution, the CNN also operates as an ICS identifier that sets thresholds to suppress ICS events. With 98 % accuracy in photoelectric absorption events, we confirm a small rod phantom, validated to have similar timing and imaging performance as for Anger logic, provides a 1.7-fold improvement in sensitivity. This CNN-based crystal identifier eliminates the need for recalibration while maintaining image quality and offering applications in maintenance, replacement, and mass production. |
