Determination of scintillation pixel location through deep learning using a two-layer DOI detector

Abstract

Small gantries and long, thin scintillation pixels are used in preclinical positron emission tomography, resulting in parallax errors outside the system’s field of view. To solve this problem, a detector for measuring the depth of interaction (DOI) was developed. In addition, conduct of research on methods for DOI measurement through deep learning is underway. In this study, we designed a detector for measurement of DOI, consisting of two layers of scintillation pixel arrays and developed a method for specifying 3-dimensional (3D) position through deep learning. DETECT2000 simulation was performed to assess the 3D-positioning accuracy of the designed detector. Data acquired through DETECT2000 simulation wereused for learning a deep learning model, and assessment of location specification accuracy was performed using data generated at a new location and the deep learning model. According to the result, the 3D-position measurement accuracy was calculated as 94.48% on average.