Modeling the Performance of Gamma-ray Source Detection and Identification Algorithms using Spectral Distinctiveness and Similarity Metrics

• Published in: 2023 IEEE Nuclear Science Symposium, Medical Imaging Conference and International Symposium on Room-Temperature Semiconductor Detectors (NSS MIC RTSD) p. 1
• Main Authors: Bandstra, M. S., Archer, D. E., Daughhetee, J. D., Ghawaly, J. M., Jones, A. C., Joshi, T. H. Y., Nachtsheim, A., Nicholson, A. D., Peplow, D. E., Prins, N., Quiter, B. J.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.11.2023
• Subjects: Gamma-rays; Machine learning; Microwave integrated circuits; Radioactive materials; Semiconductor detectors; Semiconductor device measurement; Time series analysis
• ISSN: 2577-0829
• DOI: 10.1109/NSSMICRTSD49126.2023.10338489

Over the last two decades, the widespread deployment of gamma-ray spectrometers for radiological and nuclear security applications has led to the need to develop sophisticated methods for detecting and identifying threat sources and other non-background anomalies. A major challenge has been to achieve a high probability of detection and correct identification for potentially dozens of possible source types while simultaneously maintaining operationally relevant false alarm rates, often in the presence of complex backgrounds that vary due to differing concentrations of naturally occurring radioactive material in the measurement environment. In order to help the community further develop algorithms and compare their performance, synthetic urban search datasets are being generated by the Radiological Anomaly Detection And Identification (RADAI) and RADAI-extended (REX) projects. Along with these datasets, tools are being developed to evaluate and understand the performance of algorithms on the spectral detection and identification task, focusing for now on approaches that examine single spectra rather than time series data. One of these tools is a model algorithm that performs detection and identification via template matching using various approximations and assumptions. Embedded in this model algorithm are two useful quantities that provide insight into the reasons for the absolute and relative difficulty of different source types: a factor that describes how distinguishable a source is from the background, and a factor that describes how distinguishable each source is from all the other source types. In this work we derive the model algorithm and then estimate its detection and identification performance, finally comparing those estimates to the performance of various algorithms, including machine-learning algorithms trained and evaluated using the same RADAI dataset.