Large-area radiation imaging is crucial in enhancing nuclear safety and security. Traditional mapping methods involve stationary gamma-ray detectors or imagers and necessitate multiple time-consuming static measurements at various locations.
To overcome this limitation, researchers with NSD’s Applied Nuclear Physics (ANP) program have previously introduced a new radiation imaging concept named Scene Data Fusion (SDF) [1] which enables free-moving, 3-D radiation imaging by combining conventional radiation imaging systems with contextual sensors such as video and LiDAR. Researchers have recently demonstrated new, advanced algorithms to allow the SDF concept to be more readily applied to imaging and mapping with single detectors rather than gamma-ray imagers. In single-detector mapping, variations in count rate data recorded from a free-moving detector are reconstructed into a radioactivity map. Although the concept is promising, the lack of reliable image reconstruction algorithms had previously been a limiting factor, as the conventional maximum likelihood estimation (MLE) based image reconstruction approaches are prone to overfitting.
To tackle this problem, ANP researchers, in collaboration with Jaewon Lee, a graduate student at UC Berkeley who has been leading this work, have recently developed a novel Bayesian image reconstruction algorithm based on Gaussian process priors. The approach can combat the overfitting problem of the MLE, and provide Bayesian uncertainty quantification capabilities using Bayesian posterior approximation techniques, such as the Markov Chain Monte Carlo (MCMC) algorithm and the Laplace approximation. Uncertainty quantification provides highly useful information for a variety of radiological search activities.
Furthermore, the algorithm can be extended to decompose a time series of radiation spectra collected from a single detector into constituent background and source spectral components, along with the spatial distributions of these components. This approach was demonstrated with data collected at the Savannah River National Laboratory (SRNL) H-area, where an unknown amount of 137Cs contamination is present. Despite the very low 137Cs count rates recorded (<10 cps) compared to the background (~400 cps), the algorithm successfully recovered a deposition map, along with the corresponding 137Cs and background spectral features. Additionally, the uncertainties were measured in the reconstructed deposition map and the spectral components, providing valuable information for future measurement planning. The reconstructed 137Cs activity and uncertainty maps are shown in Fig. 1 and the decomposed spectral components are shown in Fig. 2.
The ANP plans to broaden the algorithm’s application to encompass a wide array of radiation imaging problems, where uncertainty quantification is indispensable, such as Targeted Alpha Therapy (TAT) radiopharmaceutical imaging and radiation dose mapping inside the Chernobyl nuclear power plant’s new safe confinement. This promises a significant advancement in the field of radiation imaging.
Figure 1: (Left) The reconstructed image of Cs-137 activity distribution at the SRNL H-area. The measurement path along with the LiDAR point clouds are superimposed on the reconstructed activity map. Notice that the radiation background contribution changes dynamically, ranging from 200 to 500 cps. (Right) The corresponding pixelwise Bayesian uncertainty map (90% credible intervals) from the Laplace approximation.
Figure 2: (Top) The source spectral component reconstructed from the measurement at the SRNL H-area. Notice the photopeak at 662 keV and the Compton edge at 477 keV, hallmarks of a Cs-137 gamma-ray spectrum. The shaded region represents the 90% Bayesian credible intervals. (Bottom) The background spectral components recovered from the measurement. The dynamically changing background contribution can be effectively modeled by the linear combination of the three background components recovered.
References
[1] D. Hellfeld, et al, “Free-moving quantitative gamma-ray imaging,” Sci. Rep. 11 20515 (2021), https://doi.org/10.1038/s41598-021-99588-z