Convolutional Neural Network for Crystal Identification and Gamma Ray Localization in PET

Spatial resolution in positron emission tomography using traditional pixelated block detectors is inherently limited by the size of the detector array elements, namely, the scintillator crystals and readout pixels. Conventional centroiding algorithms based on Anger logic are widely used to localize individual events down to the scintillator level. However, these algorithms are associated with well-known performance degradation along the edges and corners of detector arrays. In this article, we explore the use of convolutional neural networks (CNNs) for 3D gamma ray localization as a computationally inexpensive alternative to classical centroiding. The method is successfully implemented on Monte Carlo simulated data from a single-ended readout depth-encoding detector array consisting of LYSO:Ce scintillator crystals coupled 4-to-1 to silicon photomultiplier (SiPM) pixels. The CNN demonstrated higher crystal identification accuracy at the edges and corners (99.0% versus 49.2%) and lower spatial error compared to classical centroiding (0.38 mm versus 0.76 mm). In addition, the CNN achieved 2.75 mm FWHM depth-of-interaction (DOI) resolution. Preliminary qualitative results for how our approach translates to experimental data after training on simulated data are also presented. Future work on the CNN-based approach with more experimental data could improve the performance of block detectors with multicrystal scintillators and possibly achieve subscintillator spatial resolution.