Hybrid convolutional neural network approach for optimizing automatic identification of natural isotopes in gamma ray environmental sample spectra

Abstract

Radioisotope identification presents challenges that can be effectively addressed through pattern recognition and machine learning (ML) techniques. However, further investigation is necessary to assess the accuracy of these algorithms in quantifying mixtures of radioisotopes. The novelty of the study focuses on a hybrid convolutional neural network (CNN) architecture, called Arch, which utilizes numerical values to predict the presence of radioisotopes based on their signals. The feature extraction methods are employed to analyze small-isotope libraries using area-of-interest techniques and low-resolution spectrometers, with fully calibrated detectors ensuring accurate identification complexity. Additionally, this study explores the use of two sets of machine learning approaches for the automated identification of radioisotopes, focusing specifically on the feature extraction method. The Hybrid CNN Arc model, as proposed, achieved a test data accuracy of 95%. Additionally, a recurrent neural network model achieved an accuracy of 92%, while a GBDT model achieved an accuracy of 86%. The precision, recall, and f1-score metrics have been computed for the Hybrid CNN Arch approach, yielding values of 95%, 95%, and 95%, respectively. Similarly, the RNN model achieved precision, recall, and f1-score scores of 89%, 82%, and 81.5%, respectively. Lastly, the GBDT model attained precision, recall, and f1-score values of 84%, 81%, and 74.6%, respectively.