Neural-Network-Based Energy Calculation for Multivoltage Threshold Sampling

Abstract

Multivoltage threshold (MVT) is an all digital scintillation pulse sampling method with advantages over traditional analog–digital mixed approaches in positron emission tomography (PET). However, MVT has limitations in energy calculation over the whole energy spectrum. We constructed a convolutional neural network (CNN) to establish an accurate mapping between MVT samples and energy values. We acquired pulse signals from five single-channel scintillation detectors with an oscilloscope. The CNN was trained with one detector and tested on all five detectors. It was compared with integration, CNN, and the biexponential fitting methods, with respect to the energy resolution (ER), mean absolute error, and R-squared. The CNN method was also preliminarily tested on a system with 12 672 channels. The results showed that the CNN method greatly outperformed the biexponential fitting and was comparable to integration. On average, the ER was improved from 21.0%@511 keV by biexponential fitting to 12.3%@511 keV by CNN on the signal channel level, and from 24.7%@511 keV to 20.3%@511 keV on the system level. The CNN also obtained energy spectra almost identical to those of the integration method.