Radioisotope Identification with Scintillation Detector Based on Artificial Neural Networks Using Simulated Training Data

Artificial neural networks (ANN) based on learning the features of the entire measured gamma energy spectrum has been used for radioisotope identification and proved promising especially for gamma-ray spectroscopy with low energy resolution. The implementation of ANN method, however, requires tedious experimental measurement process in generation of training data for various radioisotopes. In this work, we propose an ANN-based radioisotope identification method with simulated training data. Gamma energy spectra of 27 different radioisotopes were generated with Monte Carlo simulation. A detector energy response model was proposed to match the energy spectra generated from simulation and measured from experiment, thus “pseudo” measured energy spectra of various radioisotopes transformed from simulation can be used for ANN training, which eliminates the tedious experimental measurement process for training data generation. To reduce the complexity of the training process, the principal component analysis (PCA) method was used for dimension reduction of the input energy spectra in ANN and the channel number of the energy spectra was reduced from 2000 to 50. The trained ANN was further used to identify experimentally measured gamma energy spectra of various radioisotopes including 60Co., 137CS., 18F, 131I, 226Ra and 232Th at 103, 104 and 105 count levels. In single isotope identification test, with increased count level, higher correct identification rate is achieved and at 105 count level, all the isotopes are correctly identified for all the samples. In mixed isotope identification test, at 105 count level, all the radioisotope combinations can be identified with a correct identification rate larger than 98%, which demonstrates the feasibility and accuracy of the ANN method. To conclude, the proposed ANN method with simulated training data features good radioisotope identification capability with greatly simplified training data generation process and is feasible for gamma spectroscopy with relatively poor energy resolution.