Preliminary study on scintillation pixel position determination through deep learning of a two-layer DOI detector

Abstract

Small animal positron emission tomography (PET) uses very small, elongated scintillation pixels and a small gantry for excellent spatial resolution and sensitivity. However, with this imaging structure the spatial resolution deteriorates outside the field of view (FOV). This has led to the development of various depth of interaction (DOI) detectors. In this study, we developed a DOI detector consisting of a two-layer scintillation pixel array and a light guide, and applied a deep learning algorithm to determine the three-dimensional (3D) position of the detector. The accuracy of the position measurement based on the deep learning algorithm was evaluated by using a Na-22 radiation source to generate gamma rays to acquire scintillation signals. These signals were subsequently used to build the deep learning model, whereupon the 3D localization accuracy was evaluated using newly acquired signals resulting from gamma ray interaction. The overall 3D positioning accuracy was determined to be approximately 92.28 %. This result confirmed that the 3D position of the DOI detector can be determined with the deep learning algorithm, and consequently, the position of the scintillation pixel that interacted with the gamma rays in the detector—based on the signal obtained from the optical sensor— was measurable in real time.