Shallow Convolutional Neural Network for 3D Gamma Ray Localization in High Resolution PET

PET imaging inherently suffers from relatively poor spatial resolution compared with other clinical imaging modalities due to the size of the detector array elements, namely the scintillator crystals and readout pixels. Using 4-to-1 scintillator-to-readout coupling schemes, single-ended readout detector modules have been developed that simultaneously achieve depth-encoding readout to mitigate detector-based artifacts (i.e., parallax error) as well as sub-pixel spatial resolution equal to the scintillator crystal dimensions. Anger Logic-based localization schemes, such as energy weighted averaging, are commonly used to reconstruct the positions of photoelectric interaction sites within the modules, but these methods aren't robust to edge and corner performance degradation, resulting in non-uniform position-dependent spatial performance. We explore the use of convolutional neural networks to perform 3D gamma ray localization with single-ended readout depth-encoding modules using Monte Carlo simulated PET data. Our approach to localizing gamma ray interactions in PET detectors outperforms conventional high resolution DOI readout methods, making it a step toward making cost-effective single-ended readout depth-encoding detector modules practically viable.