Radioisotope Identification Using Convolutional Neural Networks for CsI(Tl) Gamma-ray Spectrometer

Abstract

Scintillator spectrometers can be manufactured in small volumes and lightweight. Therefore, it can be used as a portable radionuclide identification device. Due to its low resolution, however, radioisotope identification (RIID) performance and its application using scintillator spectrometers are limited, particularly, in situations where multiple radioisotopes (RIs) need to be resolved. If a machine learning (ML)-based RIID is utilized, however, the limitations can be overcome with the powerful resolving power of the spectrometer. ML-based RIID has shown promise in resolving mixtures of more than one RI. RIID using Convolutional Neural Networks (CNNs), one of the ML techniques, has received considerable attention and has been actively studied recently because it can identify the best feature, such as Compton continua and, photopeak, etc. in a gamma-ray spectrum. Here, a study was conducted to further develop the performance of CNNs-based RIID and a method of changing the dimension of the RI mixture spectral data as training inputs was attempted for the first time to improve the RIID performance of the CNNs model. A CsI(Tl) spectrometer was manufactured as a portable radionuclide identification device and was used to obtain the measurement data of gamma spectra with RIs. A RI mixture spectral dataset for training was produced using a synthesis process with each single RI gamma spectrum. One more dataset for training was produced by converting the one dimensional (1-D) spectral data to two dimensional (2-D) data in the way we proposed for the purpose of extending the inceptive field and improving RIID performance of the model. After the structure and hyper-parameters of the model were determined, the CNNs model was trained with the prepared data. The RIID performance of the CNNs model trained with 2-D data was compared with that of the model trained with 1-D data. The performance of each trained model was evaluated by mean magnitude of relative error (MMRE) and the ability to predict the zero activity of each RI as performance evaluation metrics. The model trained with the transformed 2-D data outperformed the model trained with the 1-D data in RIID over the test value range of 0.05–0.4. The MMRE averages for all RIs of the model trained with 2-D data and the model trained using the 1-D data were 2.94% and 3.94%, respectively. This meant that the CNNs model trained with 2-D data was advantageous for identifying and quantifying RIs with less than 0.4 relative radioactivity. In the test range of more than 0.5, the MMRE averages for all RIs of both models were comparable, with MMRE of less than 4% when the models were further trained with additional dataset. The abilities of the two trained models to determine the presence or absence of RI were comparable. Both models can predict the relative radioactivity value of 0 for each RI to be less than 0.05 with a probability of approximately 97%. The 2-D trained model showed better RIID performance because the inceptive field was extended by the proposed input transformation and using 2-D convolution. Therefore, the model trained with the transformed 2-D could find the meaningful relationships between spatial features. The proposed method of transforming 1-D spectral data into 2-D data is advantageous for identifying and quantifying RIs, especially, in the situation where the multiple RIs or peaks need to be resolved, and is a promising approach for CNNs-based RIID of a scintillator spectrometer.