Study on Rapid Identification of Beta Radionuclides Using Artificial Intelligence Models

Abstract

When beta particles accumulate within the body, they can damage human cells and have prolonged adverse effects. Therefore, monitoring beta radionuclides is essential to ensure worker safety from beta particle exposure. Since beta radionuclides have a continuous spectrum without distinct peaks, identifying specific radionuclides by peak energy is challenging. As a result, beta radionuclide analysis typically requires pre-treatment, such as chemical separation. Liquid scintillation counters are often used for precise beta radionuclide measurements, but they are time-consuming and not suitable for real- time field measurements. To address this issue, a study was conducted on the rapid identification of beta radionuclides using artificial intelligence models. This approach involved assessing the artificial intelligence models' capability to identify beta radionuclides by using smear paper with absorbed liquid beta sources. The sources used for model performance evaluation included the beta-emitting radionuclides 60Co, 90Sr/90Y, 137Cs, and 152Eu, combined in various ways to form 15 different combinations. The artificial intelligence models applied were support vector machine (SVM) and time series classification with a transformer (TSCT), and multiple preprocessing steps were conducted to improve accuracy: hyperparameter tuning, data smoothing, and power transformation. With these additional preprocessing steps, both models achieved 100% accuracy. To explore the limitations of smear paper applicability, additional evaluations were performed to determine the lowest activity level at which identification remains feasible. The radionuclides 60Co, 90Sr/90Y, and 137Cs were accurately identified even at surface contamination levels of 20.88 Bq/cm2, 3.19 Bq/cm2, and 6.48 Bq/cm2, respectively. Mixed radionuclides were also tested with varying activity ratios, showing that when there was a threefold difference or more, the radionuclide with the higher activity was identified. When the combinations were narrowed to 10, the models achieved 100% identification accuracy across all samples. Training with the SVM and TSCT models took approximately 10 seconds and 1 minute, respectively, and testing on new data took only a few seconds.