Accurate estimation of depth of interaction in PET on monolithic crystal coupled to SiPMs using a deep neural network and Monte Carlo simulations

摘要

In PET, the depth of interaction (DOI) information is embedded in the scintillation light distribution sampled by the photodiode array. The block detector can be considered as a non-linear function projecting a beam position coordinate onto a set of photodetector (APD, SiPM) signals. The goal of positioning algorithms is to inverse this mapping and to match the set of photodetector responses with an incident event position coordinate. Furthermore, continuous crystals were investigated as an alternative to pixelated scintillator arrays in positron emission tomography (PET). Monoliths provide good energy, timing and spatial resolution, including intrinsic depth of interaction (DOI) encoding. We propose a method for estimation of the DOI using a deep neural network based on a supervised algorithm. The network was evaluated by Monte Carlo (MC) simulations of a preclinical PET scanner with ten 50×50×10 mm³ monolithic LYSO crystals and 12×12 SiPM array. All physical phenomena, especially optical interactions, were taken into consideration. The three-dimensional interaction position in each crystal was estimated by the neural network whose inputs were the detection positions on the photodetector plane (X-Y plane) and the deposited energy. Training and validation datasets were generated by the GEANT4 MC toolkit through varying single photons incident direction, angle and energy and readout of the SiPMs output. We used a multilayer perceptron with 4 layers and 256 units as neural network architecture. The optimized layers and units were optimized after comparing several architectures. The spatial resolution in the X-Y plane and Z axis (depth of interaction) were 1.54 and 1.59 mm, respectively. Furthermore, our model was able to predict the DOI below 7 mm depth with a bias under 8.7%. The proposed method enabled higher accuracy of the interaction position estimation than existing methods based on the Anger method. Therefore, estimation of the 3D interaction position based on monolithic detectors is possible using deep neural networks.