Deep Learning-Based Method for Activity Estimation From Short-Duration Gamma Spectroscopy Recordings

Gamma spectroscopy is a technique for identifying radioactive sources and evaluating their activity. Reliable estimation of activity levels remains challenging due to various physical effects, such as the pile-up effect and dead time. In numerous applications, it is critical to provide an accurate estimation of the source activity from short-duration signals. This study compares three deep neural network (DNN) architectures: Unet, long short-term memory (LSTM), and an application-tailored convolutional neural network (CNN). Due to the lack of annotated datasets in the field, experiments are performed on simulated yet realistic signals and demonstrate that the application-tailored CNN outperforms both Unet and LSTM architectures. We show that the provided activity estimators have a lower variance than current state-of-the-art methods, which makes the approach well-adapted to analyze short-time signals. However, results on real signals illustrate the necessity of fully simulating pulses to match real pulse shapes, indicating that current approaches still face challenges in practical applications for a large class of detectors.